[Script Info]
Title: 
[Events]
Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
Dialogue: 0,0:00:00.88,0:00:04.89,Default,,0000,0000,0000,,It used to be that if you wanted\Nto get a computer to do something new,
Dialogue: 0,0:00:04.89,0:00:06.45,Default,,0000,0000,0000,,you would have to program it.
Dialogue: 0,0:00:06.45,0:00:09.86,Default,,0000,0000,0000,,Now, programming, for those of you here\Nthat haven't done it yourself,
Dialogue: 0,0:00:09.86,0:00:13.36,Default,,0000,0000,0000,,requires laying out in excruciating detail
Dialogue: 0,0:00:13.36,0:00:16.73,Default,,0000,0000,0000,,every single step that you want\Nthe computer to do
Dialogue: 0,0:00:16.73,0:00:19.09,Default,,0000,0000,0000,,in order to achieve your goal.
Dialogue: 0,0:00:19.09,0:00:22.58,Default,,0000,0000,0000,,Now, if you want to do something\Nthat you don't know how to do yourself,
Dialogue: 0,0:00:22.58,0:00:24.65,Default,,0000,0000,0000,,then this is going\Nto be a great challenge.
Dialogue: 0,0:00:24.65,0:00:28.13,Default,,0000,0000,0000,,So this was the challenge faced\Nby this man, Arthur Samuel.
Dialogue: 0,0:00:28.13,0:00:32.21,Default,,0000,0000,0000,,In 1956, he wanted to get this computer
Dialogue: 0,0:00:32.21,0:00:34.55,Default,,0000,0000,0000,,to be able to beat him at checkers.
Dialogue: 0,0:00:34.55,0:00:36.59,Default,,0000,0000,0000,,How can you write a program,
Dialogue: 0,0:00:36.59,0:00:40.39,Default,,0000,0000,0000,,lay out in excruciating detail,\Nhow to be better than you at checkers?
Dialogue: 0,0:00:40.39,0:00:42.12,Default,,0000,0000,0000,,So he came up with an idea:
Dialogue: 0,0:00:42.12,0:00:45.84,Default,,0000,0000,0000,,he had the computer play\Nagainst itself thousands of times
Dialogue: 0,0:00:45.84,0:00:48.36,Default,,0000,0000,0000,,and learn how to play checkers.
Dialogue: 0,0:00:48.36,0:00:51.54,Default,,0000,0000,0000,,And indeed it worked,\Nand in fact, by 1962,
Dialogue: 0,0:00:51.54,0:00:55.56,Default,,0000,0000,0000,,this computer had beaten\Nthe Connecticut state champion.
Dialogue: 0,0:00:55.56,0:00:58.53,Default,,0000,0000,0000,,So Arthur Samuel was\Nthe father of machine learning,
Dialogue: 0,0:00:58.53,0:01:00.25,Default,,0000,0000,0000,,and I have a great debt to him,
Dialogue: 0,0:01:00.25,0:01:03.01,Default,,0000,0000,0000,,because I am a machine\Nlearning practitioner.
Dialogue: 0,0:01:03.01,0:01:04.48,Default,,0000,0000,0000,,I was the president of Kaggle,
Dialogue: 0,0:01:04.48,0:01:07.87,Default,,0000,0000,0000,,a community of over 200,000\Nmachine learning practictioners.
Dialogue: 0,0:01:07.87,0:01:09.92,Default,,0000,0000,0000,,Kaggle puts up competitions
Dialogue: 0,0:01:09.92,0:01:13.63,Default,,0000,0000,0000,,to try and get them to solve\Npreviously unsolved problems,
Dialogue: 0,0:01:13.63,0:01:17.47,Default,,0000,0000,0000,,and it's been successful \Nhundreds of times.
Dialogue: 0,0:01:17.47,0:01:19.94,Default,,0000,0000,0000,,So from this vantage point,\NI was able to find out
Dialogue: 0,0:01:19.94,0:01:23.89,Default,,0000,0000,0000,,a lot about what machine learning\Ncan do in the past, can do today,
Dialogue: 0,0:01:23.89,0:01:26.25,Default,,0000,0000,0000,,and what it could do in the future.
Dialogue: 0,0:01:26.25,0:01:30.68,Default,,0000,0000,0000,,Perhaps the first big success of \Nmachine learning commercially was Google.
Dialogue: 0,0:01:30.68,0:01:33.78,Default,,0000,0000,0000,,Google showed that it is\Npossible to find information
Dialogue: 0,0:01:33.78,0:01:35.54,Default,,0000,0000,0000,,by using a computer algorithm,
Dialogue: 0,0:01:35.54,0:01:38.44,Default,,0000,0000,0000,,and this algorithm is based\Non machine learning.
Dialogue: 0,0:01:38.44,0:01:42.32,Default,,0000,0000,0000,,Since that time, there have been many\Ncommercial successes of machine learning.
Dialogue: 0,0:01:42.32,0:01:44.16,Default,,0000,0000,0000,,Companies like Amazon and Netflix
Dialogue: 0,0:01:44.16,0:01:47.88,Default,,0000,0000,0000,,use machine learning to suggest\Nproducts that you might like to buy,
Dialogue: 0,0:01:47.88,0:01:49.90,Default,,0000,0000,0000,,movies that you might like to watch.
Dialogue: 0,0:01:49.90,0:01:51.70,Default,,0000,0000,0000,,Sometimes, it's almost creepy.
Dialogue: 0,0:01:51.70,0:01:53.66,Default,,0000,0000,0000,,Companies like LinkedIn and Facebook
Dialogue: 0,0:01:53.66,0:01:56.25,Default,,0000,0000,0000,,sometimes will tell you about\Nwho your friends might be
Dialogue: 0,0:01:56.25,0:01:58.23,Default,,0000,0000,0000,,and you have no idea how it did it,
Dialogue: 0,0:01:58.23,0:02:01.20,Default,,0000,0000,0000,,and this is because it's using\Nthe power of machine learning.
Dialogue: 0,0:02:01.20,0:02:04.15,Default,,0000,0000,0000,,These are algorithms that have\Nlearned how to do this from data
Dialogue: 0,0:02:04.15,0:02:07.40,Default,,0000,0000,0000,,rather than being programmed by hand.
Dialogue: 0,0:02:07.40,0:02:09.88,Default,,0000,0000,0000,,This is also how IBM was successful
Dialogue: 0,0:02:09.88,0:02:13.74,Default,,0000,0000,0000,,in getting Watson to beat\Nthe two world champions at "Jeopardy,"
Dialogue: 0,0:02:13.74,0:02:16.96,Default,,0000,0000,0000,,answering incredibly subtle\Nand complex questions like this one.
Dialogue: 0,0:02:16.96,0:02:19.80,Default,,0000,0000,0000,,["The ancient 'Lion of Nimrud' went missing\Nfrom this city's national museum in 2003 \N(along with a lot of other stuff)"]
Dialogue: 0,0:02:19.80,0:02:23.03,Default,,0000,0000,0000,,This is also why we are now able\Nto see the first self-driving cars.
Dialogue: 0,0:02:23.03,0:02:25.86,Default,,0000,0000,0000,,If you want to be able to tell\Nthe difference between, say,
Dialogue: 0,0:02:25.86,0:02:28.49,Default,,0000,0000,0000,,a tree and a pedestrian,\Nwell, that's pretty important.
Dialogue: 0,0:02:28.49,0:02:31.08,Default,,0000,0000,0000,,We don't know how to write\Nthose programs by hand,
Dialogue: 0,0:02:31.08,0:02:34.07,Default,,0000,0000,0000,,but with machine learning,\Nthis is now possible.
Dialogue: 0,0:02:34.07,0:02:36.68,Default,,0000,0000,0000,,And in fact, this car has driven \Nover a million miles
Dialogue: 0,0:02:36.68,0:02:40.19,Default,,0000,0000,0000,,without any accidents on regular roads.
Dialogue: 0,0:02:40.20,0:02:44.11,Default,,0000,0000,0000,,So we now know that computers can learn,
Dialogue: 0,0:02:44.11,0:02:46.01,Default,,0000,0000,0000,,and computers can learn to do things
Dialogue: 0,0:02:46.01,0:02:48.85,Default,,0000,0000,0000,,that we actually sometimes\Ndon't know how to do ourselves,
Dialogue: 0,0:02:48.85,0:02:51.73,Default,,0000,0000,0000,,or maybe can do them better than us.
Dialogue: 0,0:02:51.73,0:02:55.93,Default,,0000,0000,0000,,One of the most amazing examples\NI've seen of machine learning
Dialogue: 0,0:02:55.93,0:02:58.32,Default,,0000,0000,0000,,happened on a project that I ran at Kaggle
Dialogue: 0,0:02:58.32,0:03:01.91,Default,,0000,0000,0000,,where a team run by a guy\Ncalled Geoffrey Hinton
Dialogue: 0,0:03:01.91,0:03:03.46,Default,,0000,0000,0000,,from the University of Toronto
Dialogue: 0,0:03:03.46,0:03:06.14,Default,,0000,0000,0000,,won a competition for\Nautomatic drug discovery.
Dialogue: 0,0:03:06.14,0:03:08.99,Default,,0000,0000,0000,,Now, what was extraordinary here\Nis not just that they beat
Dialogue: 0,0:03:08.99,0:03:13.00,Default,,0000,0000,0000,,all of the algorithms developed by Merck\Nor the international academic community,
Dialogue: 0,0:03:13.00,0:03:18.06,Default,,0000,0000,0000,,but nobody on the team had any background\Nin chemistry or biology or life sciences,
Dialogue: 0,0:03:18.06,0:03:20.23,Default,,0000,0000,0000,,and they did it in two weeks.
Dialogue: 0,0:03:20.23,0:03:21.61,Default,,0000,0000,0000,,How did they do this?
Dialogue: 0,0:03:22.42,0:03:25.34,Default,,0000,0000,0000,,They used an extraordinary algorithm\Ncalled deep learning.
Dialogue: 0,0:03:25.34,0:03:28.29,Default,,0000,0000,0000,,So important was this that in fact\Nthe success was covered
Dialogue: 0,0:03:28.29,0:03:31.41,Default,,0000,0000,0000,,in The New York Times in a front page\Narticle a few weeks later.
Dialogue: 0,0:03:31.41,0:03:34.15,Default,,0000,0000,0000,,This is Geoffrey Hinton\Nhere on the left-hand side.
Dialogue: 0,0:03:34.15,0:03:38.49,Default,,0000,0000,0000,,Deep learning is an algorithm\Ninspired by how the human brain works,
Dialogue: 0,0:03:38.49,0:03:40.30,Default,,0000,0000,0000,,and as a result it's an algorithm
Dialogue: 0,0:03:40.30,0:03:44.14,Default,,0000,0000,0000,,which has no theoretical limitations\Non what it can do.
Dialogue: 0,0:03:44.14,0:03:46.96,Default,,0000,0000,0000,,The more data you give it and the more\Ncomputation time you give it,
Dialogue: 0,0:03:46.96,0:03:48.28,Default,,0000,0000,0000,,the better it gets.
Dialogue: 0,0:03:48.28,0:03:50.62,Default,,0000,0000,0000,,The New York Times also\Nshowed in this article
Dialogue: 0,0:03:50.62,0:03:52.86,Default,,0000,0000,0000,,another extraordinary\Nresult of deep learning
Dialogue: 0,0:03:52.86,0:03:55.57,Default,,0000,0000,0000,,which I'm going to show you now.
Dialogue: 0,0:03:55.57,0:04:00.51,Default,,0000,0000,0000,,It shows that computers \Ncan listen and understand.
Dialogue: 0,0:04:00.51,0:04:03.22,Default,,0000,0000,0000,,(Video) Richard Rashid: Now, the last step
Dialogue: 0,0:04:03.22,0:04:06.25,Default,,0000,0000,0000,,that I want to be able\Nto take in this process
Dialogue: 0,0:04:06.25,0:04:10.96,Default,,0000,0000,0000,,is to actually speak to you in Chinese.
Dialogue: 0,0:04:10.96,0:04:13.60,Default,,0000,0000,0000,,Now the key thing there is,
Dialogue: 0,0:04:13.60,0:04:18.60,Default,,0000,0000,0000,,we've been able to take a large amount \Nof information from many Chinese speakers
Dialogue: 0,0:04:18.60,0:04:21.13,Default,,0000,0000,0000,,and produce a text-to-speech system
Dialogue: 0,0:04:21.13,0:04:25.80,Default,,0000,0000,0000,,that takes Chinese text\Nand converts it into Chinese language,
Dialogue: 0,0:04:25.80,0:04:29.93,Default,,0000,0000,0000,,and then we've taken\Nan hour or so of my own voice
Dialogue: 0,0:04:29.93,0:04:31.82,Default,,0000,0000,0000,,and we've used that to modulate
Dialogue: 0,0:04:31.82,0:04:36.36,Default,,0000,0000,0000,,the standard text-to-speech system\Nso that it would sound like me.
Dialogue: 0,0:04:36.36,0:04:38.90,Default,,0000,0000,0000,,Again, the result's not perfect.
Dialogue: 0,0:04:38.90,0:04:41.55,Default,,0000,0000,0000,,There are in fact quite a few errors.
Dialogue: 0,0:04:41.55,0:04:44.04,Default,,0000,0000,0000,,(In Chinese)
Dialogue: 0,0:04:44.04,0:04:47.40,Default,,0000,0000,0000,,(Applause)
Dialogue: 0,0:04:49.45,0:04:53.02,Default,,0000,0000,0000,,There's much work to be done in this area.
Dialogue: 0,0:04:53.02,0:04:56.67,Default,,0000,0000,0000,,(In Chinese)
Dialogue: 0,0:04:56.67,0:05:00.10,Default,,0000,0000,0000,,(Applause)
Dialogue: 0,0:05:01.34,0:05:04.74,Default,,0000,0000,0000,,Jeremy Howard: Well, that was at\Na machine learning conference in China.
Dialogue: 0,0:05:04.74,0:05:07.11,Default,,0000,0000,0000,,It's not often, actually,\Nat academic conferences
Dialogue: 0,0:05:07.11,0:05:09.01,Default,,0000,0000,0000,,that you do hear spontaneous applause,
Dialogue: 0,0:05:09.01,0:05:12.69,Default,,0000,0000,0000,,although of course sometimes\Nat TEDx conferences, feel free.
Dialogue: 0,0:05:12.69,0:05:15.48,Default,,0000,0000,0000,,Everything you saw there\Nwas happening with deep learning.
Dialogue: 0,0:05:15.48,0:05:17.01,Default,,0000,0000,0000,,(Applause) Thank you.
Dialogue: 0,0:05:17.01,0:05:19.29,Default,,0000,0000,0000,,The transcription in English\Nwas deep learning.
Dialogue: 0,0:05:19.29,0:05:22.70,Default,,0000,0000,0000,,The translation to Chinese and the text\Nin the top right, deep learning,
Dialogue: 0,0:05:22.70,0:05:26.01,Default,,0000,0000,0000,,and the construction of the voice\Nwas deep learning as well.
Dialogue: 0,0:05:26.01,0:05:29.24,Default,,0000,0000,0000,,So deep learning is\Nthis extraordinary thing.
Dialogue: 0,0:05:29.24,0:05:32.34,Default,,0000,0000,0000,,It's a single algorithm that\Ncan seem to do almost anything,
Dialogue: 0,0:05:32.34,0:05:35.45,Default,,0000,0000,0000,,and I discovered that a year earlier,\Nit had also learned to see.
Dialogue: 0,0:05:35.45,0:05:37.63,Default,,0000,0000,0000,,In this obscure competition from Germany
Dialogue: 0,0:05:37.63,0:05:40.22,Default,,0000,0000,0000,,called the German Traffic Sign \NRecognition Benchmark,
Dialogue: 0,0:05:40.22,0:05:43.62,Default,,0000,0000,0000,,deep learning had learned\Nto recognize traffic signs like this one.
Dialogue: 0,0:05:43.62,0:05:45.71,Default,,0000,0000,0000,,Not only could it\Nrecognize the traffic signs
Dialogue: 0,0:05:45.71,0:05:47.47,Default,,0000,0000,0000,,better than any other algorithm,
Dialogue: 0,0:05:47.47,0:05:50.19,Default,,0000,0000,0000,,the leaderboard actually showed\Nit was better than people,
Dialogue: 0,0:05:50.19,0:05:52.04,Default,,0000,0000,0000,,about twice as good as people.
Dialogue: 0,0:05:52.04,0:05:54.04,Default,,0000,0000,0000,,So by 2011, we had the first example
Dialogue: 0,0:05:54.04,0:05:57.44,Default,,0000,0000,0000,,of computers that can see\Nbetter than people.
Dialogue: 0,0:05:57.44,0:05:59.49,Default,,0000,0000,0000,,Since that time, a lot has happened.
Dialogue: 0,0:05:59.49,0:06:03.00,Default,,0000,0000,0000,,In 2012, Google announced that\Nthey had a deep learning algorithm
Dialogue: 0,0:06:03.00,0:06:04.42,Default,,0000,0000,0000,,watch YouTube videos
Dialogue: 0,0:06:04.42,0:06:07.86,Default,,0000,0000,0000,,and crunched the data\Non 16,000 computers for a month,
Dialogue: 0,0:06:07.86,0:06:12.22,Default,,0000,0000,0000,,and the computer independently learned\Nabout concepts such as people and cats
Dialogue: 0,0:06:12.22,0:06:14.03,Default,,0000,0000,0000,,just by watching the videos.
Dialogue: 0,0:06:14.03,0:06:16.38,Default,,0000,0000,0000,,This is much like the way\Nthat humans learn.
Dialogue: 0,0:06:16.38,0:06:19.12,Default,,0000,0000,0000,,Humans don't learn\Nby being told what they see,
Dialogue: 0,0:06:19.12,0:06:22.45,Default,,0000,0000,0000,,but by learning for themselves\Nwhat these things are.
Dialogue: 0,0:06:22.45,0:06:25.82,Default,,0000,0000,0000,,Also in 2012, Geoffrey Hinton,\Nwho we saw earlier,
Dialogue: 0,0:06:25.82,0:06:28.68,Default,,0000,0000,0000,,won the very popular ImageNet competition,
Dialogue: 0,0:06:28.68,0:06:32.82,Default,,0000,0000,0000,,looking to try to figure out \Nfrom one and a half million images
Dialogue: 0,0:06:32.82,0:06:34.26,Default,,0000,0000,0000,,what they're pictures of.
Dialogue: 0,0:06:34.26,0:06:37.79,Default,,0000,0000,0000,,As of 2014, we're now down\Nto a six percent error rate
Dialogue: 0,0:06:37.79,0:06:39.24,Default,,0000,0000,0000,,in image recognition.
Dialogue: 0,0:06:39.24,0:06:41.27,Default,,0000,0000,0000,,This is better than people, again.
Dialogue: 0,0:06:41.27,0:06:45.04,Default,,0000,0000,0000,,So machines really are doing\Nan extraordinarily good job of this,
Dialogue: 0,0:06:45.04,0:06:47.31,Default,,0000,0000,0000,,and it is now being used in industry.
Dialogue: 0,0:06:47.31,0:06:50.35,Default,,0000,0000,0000,,For example, Google announced last year
Dialogue: 0,0:06:50.35,0:06:54.93,Default,,0000,0000,0000,,that they had mapped every single\Nlocation in France in two hours,
Dialogue: 0,0:06:54.93,0:06:58.38,Default,,0000,0000,0000,,and the way they did it was\Nthat they fed street view images
Dialogue: 0,0:06:58.38,0:07:02.70,Default,,0000,0000,0000,,into a deep learning algorithm\Nto recognize and read street numbers.
Dialogue: 0,0:07:02.70,0:07:04.92,Default,,0000,0000,0000,,Imagine how long\Nit would have taken before:
Dialogue: 0,0:07:04.92,0:07:08.27,Default,,0000,0000,0000,,dozens of people, many years.
Dialogue: 0,0:07:08.27,0:07:10.18,Default,,0000,0000,0000,,This is also happening in China.
Dialogue: 0,0:07:10.18,0:07:14.22,Default,,0000,0000,0000,,Baidu is kind of \Nthe Chinese Google, I guess,
Dialogue: 0,0:07:14.22,0:07:16.50,Default,,0000,0000,0000,,and what you see here in the top left
Dialogue: 0,0:07:16.50,0:07:20.48,Default,,0000,0000,0000,,is an example of a picture that I uploaded\Nto Baidu's deep learning system,
Dialogue: 0,0:07:20.48,0:07:24.25,Default,,0000,0000,0000,,and underneath you can see that the system\Nhas understood what that picture is
Dialogue: 0,0:07:24.25,0:07:26.48,Default,,0000,0000,0000,,and found similar images.
Dialogue: 0,0:07:26.48,0:07:29.22,Default,,0000,0000,0000,,The similar images actually\Nhave similar backgrounds,
Dialogue: 0,0:07:29.22,0:07:30.88,Default,,0000,0000,0000,,similar directions of the faces,
Dialogue: 0,0:07:30.88,0:07:32.66,Default,,0000,0000,0000,,even some with their tongue out.
Dialogue: 0,0:07:32.66,0:07:35.70,Default,,0000,0000,0000,,This is not clearly looking\Nat the text of a web page.
Dialogue: 0,0:07:35.70,0:07:37.11,Default,,0000,0000,0000,,All I uploaded was an image.
Dialogue: 0,0:07:37.11,0:07:41.13,Default,,0000,0000,0000,,So we now have computers which\Nreally understand what they see
Dialogue: 0,0:07:41.13,0:07:42.75,Default,,0000,0000,0000,,and can therefore search databases
Dialogue: 0,0:07:42.75,0:07:46.31,Default,,0000,0000,0000,,of hundreds of millions\Nof images in real time.
Dialogue: 0,0:07:46.31,0:07:49.54,Default,,0000,0000,0000,,So what does it mean\Nnow that computers can see?
Dialogue: 0,0:07:49.54,0:07:51.55,Default,,0000,0000,0000,,Well, it's not just \Nthat computers can see.
Dialogue: 0,0:07:51.55,0:07:53.62,Default,,0000,0000,0000,,In fact, deep learning\Nhas done more than that.
Dialogue: 0,0:07:53.62,0:07:56.57,Default,,0000,0000,0000,,Complex, nuanced sentences like this one
Dialogue: 0,0:07:56.57,0:07:59.39,Default,,0000,0000,0000,,are now understandable\Nwith deep learning algorithms.
Dialogue: 0,0:07:59.39,0:08:00.70,Default,,0000,0000,0000,,As you can see here,
Dialogue: 0,0:08:00.70,0:08:03.46,Default,,0000,0000,0000,,this Stanford-based system\Nshowing the red dot at the top
Dialogue: 0,0:08:03.46,0:08:07.38,Default,,0000,0000,0000,,has figured out that this sentence\Nis expressing negative sentiment.
Dialogue: 0,0:08:07.38,0:08:10.79,Default,,0000,0000,0000,,Deep learning now in fact\Nis near human performance
Dialogue: 0,0:08:10.80,0:08:15.92,Default,,0000,0000,0000,,at understanding what sentences are about\Nand what it is saying about those things.
Dialogue: 0,0:08:15.92,0:08:18.65,Default,,0000,0000,0000,,Also, deep learning has\Nbeen used to read Chinese,
Dialogue: 0,0:08:18.65,0:08:21.81,Default,,0000,0000,0000,,again at about native\NChinese speaker level.
Dialogue: 0,0:08:21.81,0:08:23.98,Default,,0000,0000,0000,,This algorithm developed\Nout of Switzerland
Dialogue: 0,0:08:23.98,0:08:27.33,Default,,0000,0000,0000,,by people, none of whom speak\Nor understand any Chinese.
Dialogue: 0,0:08:27.33,0:08:29.38,Default,,0000,0000,0000,,As I say, using deep learning
Dialogue: 0,0:08:29.38,0:08:31.60,Default,,0000,0000,0000,,is about the best system\Nin the world for this,
Dialogue: 0,0:08:31.60,0:08:36.72,Default,,0000,0000,0000,,even compared to native\Nhuman understanding.
Dialogue: 0,0:08:36.72,0:08:39.68,Default,,0000,0000,0000,,This is a system that we\Nput together at my company
Dialogue: 0,0:08:39.68,0:08:41.73,Default,,0000,0000,0000,,which shows putting\Nall this stuff together.
Dialogue: 0,0:08:41.73,0:08:44.19,Default,,0000,0000,0000,,These are pictures which\Nhave no text attached,
Dialogue: 0,0:08:44.19,0:08:46.54,Default,,0000,0000,0000,,and as I'm typing in here sentences,
Dialogue: 0,0:08:46.54,0:08:49.51,Default,,0000,0000,0000,,in real time it's understanding\Nthese pictures
Dialogue: 0,0:08:49.51,0:08:51.19,Default,,0000,0000,0000,,and figuring out what they're about
Dialogue: 0,0:08:51.19,0:08:54.35,Default,,0000,0000,0000,,and finding pictures that are similar\Nto the text that I'm writing.
Dialogue: 0,0:08:54.35,0:08:57.11,Default,,0000,0000,0000,,So you can see, it's actually\Nunderstanding my sentences
Dialogue: 0,0:08:57.11,0:08:59.33,Default,,0000,0000,0000,,and actually understanding these pictures.
Dialogue: 0,0:08:59.33,0:09:01.89,Default,,0000,0000,0000,,I know that you've seen\Nsomething like this on Google,
Dialogue: 0,0:09:01.89,0:09:04.67,Default,,0000,0000,0000,,where you can type in things\Nand it will show you pictures,
Dialogue: 0,0:09:04.67,0:09:08.09,Default,,0000,0000,0000,,but actually what it's doing is it's\Nsearching the webpage for the text.
Dialogue: 0,0:09:08.09,0:09:11.09,Default,,0000,0000,0000,,This is very different from actually\Nunderstanding the images.
Dialogue: 0,0:09:11.09,0:09:13.84,Default,,0000,0000,0000,,This is something that computers\Nhave only been able to do
Dialogue: 0,0:09:13.84,0:09:17.09,Default,,0000,0000,0000,,for the first time in the last few months.
Dialogue: 0,0:09:17.09,0:09:21.18,Default,,0000,0000,0000,,So we can see now that computers\Ncan not only see but they can also read,
Dialogue: 0,0:09:21.18,0:09:24.95,Default,,0000,0000,0000,,and, of course, we've shown that they\Ncan understand what they hear.
Dialogue: 0,0:09:24.95,0:09:28.39,Default,,0000,0000,0000,,Perhaps not surprising now that\NI'm going to tell you they can write.
Dialogue: 0,0:09:28.39,0:09:33.17,Default,,0000,0000,0000,,Here is some text that I generated\Nusing a deep learning algorithm yesterday.
Dialogue: 0,0:09:33.17,0:09:37.10,Default,,0000,0000,0000,,And here is some text that an algorithm\Nout of Stanford generated.
Dialogue: 0,0:09:37.10,0:09:38.86,Default,,0000,0000,0000,,Each of these sentences was generated
Dialogue: 0,0:09:38.86,0:09:43.11,Default,,0000,0000,0000,,by a deep learning algorithm\Nto describe each of those pictures.
Dialogue: 0,0:09:43.11,0:09:47.58,Default,,0000,0000,0000,,This algorithm before has never seen\Na man in a black shirt playing a guitar.
Dialogue: 0,0:09:47.58,0:09:49.80,Default,,0000,0000,0000,,It's seen a man before,\Nit's seen black before,
Dialogue: 0,0:09:49.80,0:09:51.40,Default,,0000,0000,0000,,it's seen a guitar before,
Dialogue: 0,0:09:51.40,0:09:55.69,Default,,0000,0000,0000,,but it has independently generated\Nthis novel description of this picture.
Dialogue: 0,0:09:55.69,0:09:59.20,Default,,0000,0000,0000,,We're still not quite at human\Nperformance here, but we're close.
Dialogue: 0,0:09:59.20,0:10:03.26,Default,,0000,0000,0000,,In tests, humans prefer\Nthe computer-generated caption
Dialogue: 0,0:10:03.26,0:10:04.79,Default,,0000,0000,0000,,one out of four times.
Dialogue: 0,0:10:04.79,0:10:06.86,Default,,0000,0000,0000,,Now this system is now only two weeks old,
Dialogue: 0,0:10:06.86,0:10:08.70,Default,,0000,0000,0000,,so probably within the next year,
Dialogue: 0,0:10:08.70,0:10:11.50,Default,,0000,0000,0000,,the computer algorithm will be\Nwell past human performance
Dialogue: 0,0:10:11.50,0:10:13.36,Default,,0000,0000,0000,,at the rate things are going.
Dialogue: 0,0:10:13.36,0:10:16.41,Default,,0000,0000,0000,,So computers can also write.
Dialogue: 0,0:10:16.41,0:10:19.89,Default,,0000,0000,0000,,So we put all this together and it leads\Nto very exciting opportunities.
Dialogue: 0,0:10:19.89,0:10:21.38,Default,,0000,0000,0000,,For example, in medicine,
Dialogue: 0,0:10:21.38,0:10:23.90,Default,,0000,0000,0000,,a team in Boston announced\Nthat they had discovered
Dialogue: 0,0:10:23.90,0:10:26.85,Default,,0000,0000,0000,,dozens of new clinically relevant features
Dialogue: 0,0:10:26.85,0:10:31.12,Default,,0000,0000,0000,,of tumors which help doctors\Nmake a prognosis of a cancer.
Dialogue: 0,0:10:32.22,0:10:34.52,Default,,0000,0000,0000,,Very similarly, in Stanford,
Dialogue: 0,0:10:34.52,0:10:38.18,Default,,0000,0000,0000,,a group there announced that,\Nlooking at tissues under magnification,
Dialogue: 0,0:10:38.18,0:10:40.56,Default,,0000,0000,0000,,they've developed \Na machine learning-based system
Dialogue: 0,0:10:40.56,0:10:43.14,Default,,0000,0000,0000,,which in fact is better\Nthan human pathologists
Dialogue: 0,0:10:43.14,0:10:47.52,Default,,0000,0000,0000,,at predicting survival rates\Nfor cancer sufferers.
Dialogue: 0,0:10:47.52,0:10:50.76,Default,,0000,0000,0000,,In both of these cases, not only\Nwere the predictions more accurate,
Dialogue: 0,0:10:50.76,0:10:53.27,Default,,0000,0000,0000,,but they generated new insightful science.
Dialogue: 0,0:10:53.28,0:10:54.78,Default,,0000,0000,0000,,In the radiology case,
Dialogue: 0,0:10:54.78,0:10:57.88,Default,,0000,0000,0000,,they were new clinical indicators\Nthat humans can understand.
Dialogue: 0,0:10:57.88,0:10:59.67,Default,,0000,0000,0000,,In this pathology case,
Dialogue: 0,0:10:59.67,0:11:04.17,Default,,0000,0000,0000,,the computer system actually discovered\Nthat the cells around the cancer
Dialogue: 0,0:11:04.17,0:11:07.51,Default,,0000,0000,0000,,are as important as\Nthe cancer cells themselves
Dialogue: 0,0:11:07.51,0:11:09.26,Default,,0000,0000,0000,,in making a diagnosis.
Dialogue: 0,0:11:09.26,0:11:14.62,Default,,0000,0000,0000,,This is the opposite of what pathologists\Nhad been taught for decades.
Dialogue: 0,0:11:14.62,0:11:17.91,Default,,0000,0000,0000,,In each of those two cases,\Nthey were systems developed
Dialogue: 0,0:11:17.91,0:11:21.53,Default,,0000,0000,0000,,by a combination of medical experts\Nand machine learning experts,
Dialogue: 0,0:11:21.53,0:11:24.28,Default,,0000,0000,0000,,but as of last year,\Nwe're now beyond that too.
Dialogue: 0,0:11:24.28,0:11:27.82,Default,,0000,0000,0000,,This is an example of\Nidentifying cancerous areas
Dialogue: 0,0:11:27.82,0:11:30.35,Default,,0000,0000,0000,,of human tissue under a microscope.
Dialogue: 0,0:11:30.35,0:11:34.97,Default,,0000,0000,0000,,The system being shown here\Ncan identify those areas more accurately,
Dialogue: 0,0:11:34.97,0:11:37.74,Default,,0000,0000,0000,,or about as accurately,\Nas human pathologists,
Dialogue: 0,0:11:37.74,0:11:41.13,Default,,0000,0000,0000,,but was built entirely with deep learning\Nusing no medical expertise
Dialogue: 0,0:11:41.13,0:11:43.66,Default,,0000,0000,0000,,by people who have\Nno background in the field.
Dialogue: 0,0:11:44.73,0:11:47.28,Default,,0000,0000,0000,,Similarly, here, this neuron segmentation.
Dialogue: 0,0:11:47.28,0:11:50.95,Default,,0000,0000,0000,,We can now segment neurons\Nabout as accurately as humans can,
Dialogue: 0,0:11:50.95,0:11:53.67,Default,,0000,0000,0000,,but this system was developed\Nwith deep learning
Dialogue: 0,0:11:53.67,0:11:56.92,Default,,0000,0000,0000,,using people with no previous \Nbackground in medicine.
Dialogue: 0,0:11:56.92,0:12:00.15,Default,,0000,0000,0000,,So myself, as somebody with\Nno previous background in medicine,
Dialogue: 0,0:12:00.15,0:12:03.88,Default,,0000,0000,0000,,I seem to be entirely well qualified\Nto start a new medical company,
Dialogue: 0,0:12:03.88,0:12:06.02,Default,,0000,0000,0000,,which I did.
Dialogue: 0,0:12:06.02,0:12:07.76,Default,,0000,0000,0000,,I was kind of terrified of doing it,
Dialogue: 0,0:12:07.76,0:12:10.65,Default,,0000,0000,0000,,but the theory seemed to suggest\Nthat it ought to be possible
Dialogue: 0,0:12:10.65,0:12:16.14,Default,,0000,0000,0000,,to do very useful medicine\Nusing just these data analytic techniques.
Dialogue: 0,0:12:16.14,0:12:18.62,Default,,0000,0000,0000,,And thankfully, the feedback\Nhas been fantastic,
Dialogue: 0,0:12:18.62,0:12:20.98,Default,,0000,0000,0000,,not just from the media\Nbut from the medical community,
Dialogue: 0,0:12:20.98,0:12:23.32,Default,,0000,0000,0000,,who have been very supportive.
Dialogue: 0,0:12:23.32,0:12:27.47,Default,,0000,0000,0000,,The theory is that we can take\Nthe middle part of the medical process
Dialogue: 0,0:12:27.47,0:12:30.36,Default,,0000,0000,0000,,and turn that into data analysis\Nas much as possible,
Dialogue: 0,0:12:30.36,0:12:33.43,Default,,0000,0000,0000,,leaving doctors to do\Nwhat they're best at.
Dialogue: 0,0:12:33.43,0:12:35.03,Default,,0000,0000,0000,,I want to give you an example.
Dialogue: 0,0:12:35.03,0:12:39.98,Default,,0000,0000,0000,,It now takes us about 15 minutes\Nto generate a new medical diagnostic test
Dialogue: 0,0:12:39.98,0:12:41.93,Default,,0000,0000,0000,,and I'll show you that in real time now,
Dialogue: 0,0:12:41.93,0:12:45.42,Default,,0000,0000,0000,,but I've compressed it down to \Nthree minutes by cutting some pieces out.
Dialogue: 0,0:12:45.42,0:12:48.48,Default,,0000,0000,0000,,Rather than showing you\Ncreating a medical diagnostic test,
Dialogue: 0,0:12:48.48,0:12:51.85,Default,,0000,0000,0000,,I'm going to show you \Na diagnostic test of car images,
Dialogue: 0,0:12:51.85,0:12:54.07,Default,,0000,0000,0000,,because that's something\Nwe can all understand.
Dialogue: 0,0:12:54.07,0:12:57.27,Default,,0000,0000,0000,,So here we're starting with \Nabout 1.5 million car images,
Dialogue: 0,0:12:57.27,0:13:00.48,Default,,0000,0000,0000,,and I want to create something\Nthat can split them into the angle
Dialogue: 0,0:13:00.48,0:13:02.70,Default,,0000,0000,0000,,of the photo that's being taken.
Dialogue: 0,0:13:02.70,0:13:06.59,Default,,0000,0000,0000,,So these images are entirely unlabeled,\Nso I have to start from scratch.
Dialogue: 0,0:13:06.59,0:13:08.45,Default,,0000,0000,0000,,With our deep learning algorithm,
Dialogue: 0,0:13:08.45,0:13:12.16,Default,,0000,0000,0000,,it can automatically identify\Nareas of structure in these images.
Dialogue: 0,0:13:12.16,0:13:15.78,Default,,0000,0000,0000,,So the nice thing is that the human\Nand the computer can now work together.
Dialogue: 0,0:13:15.78,0:13:17.96,Default,,0000,0000,0000,,So the human, as you can see here,
Dialogue: 0,0:13:17.96,0:13:20.63,Default,,0000,0000,0000,,is telling the computer\Nabout areas of interest
Dialogue: 0,0:13:20.63,0:13:25.28,Default,,0000,0000,0000,,which it wants the computer then\Nto try and use to improve its algorithm.
Dialogue: 0,0:13:25.28,0:13:29.58,Default,,0000,0000,0000,,Now, these deep learning systems actually\Nare in 16,000-dimensional space,
Dialogue: 0,0:13:29.58,0:13:33.01,Default,,0000,0000,0000,,so you can see here the computer\Nrotating this through that space,
Dialogue: 0,0:13:33.01,0:13:35.00,Default,,0000,0000,0000,,trying to find new areas of structure.
Dialogue: 0,0:13:35.00,0:13:36.78,Default,,0000,0000,0000,,And when it does so successfully,
Dialogue: 0,0:13:36.78,0:13:40.79,Default,,0000,0000,0000,,the human who is driving it can then\Npoint out the areas that are interesting.
Dialogue: 0,0:13:40.79,0:13:43.21,Default,,0000,0000,0000,,So here, the computer has\Nsuccessfully found areas,
Dialogue: 0,0:13:43.21,0:13:45.77,Default,,0000,0000,0000,,for example, angles.
Dialogue: 0,0:13:45.77,0:13:47.38,Default,,0000,0000,0000,,So as we go through this process,
Dialogue: 0,0:13:47.38,0:13:49.72,Default,,0000,0000,0000,,we're gradually telling\Nthe computer more and more
Dialogue: 0,0:13:49.72,0:13:52.14,Default,,0000,0000,0000,,about the kinds of structures\Nwe're looking for.
Dialogue: 0,0:13:52.14,0:13:53.92,Default,,0000,0000,0000,,You can imagine in a diagnostic test
Dialogue: 0,0:13:53.92,0:13:57.27,Default,,0000,0000,0000,,this would be a pathologist identifying\Nareas of pathosis, for example,
Dialogue: 0,0:13:57.27,0:14:02.29,Default,,0000,0000,0000,,or a radiologist indicating\Npotentially troublesome nodules.
Dialogue: 0,0:14:02.29,0:14:04.85,Default,,0000,0000,0000,,And sometimes it can be\Ndifficult for the algorithm.
Dialogue: 0,0:14:04.85,0:14:06.82,Default,,0000,0000,0000,,In this case, it got kind of confused.
Dialogue: 0,0:14:06.82,0:14:09.36,Default,,0000,0000,0000,,The fronts and the backs\Nof the cars are all mixed up.
Dialogue: 0,0:14:09.36,0:14:11.44,Default,,0000,0000,0000,,So here we have to be a bit more careful,
Dialogue: 0,0:14:11.44,0:14:14.67,Default,,0000,0000,0000,,manually selecting these fronts\Nas opposed to the backs,
Dialogue: 0,0:14:14.67,0:14:20.18,Default,,0000,0000,0000,,then telling the computer\Nthat this is a type of group
Dialogue: 0,0:14:20.18,0:14:21.52,Default,,0000,0000,0000,,that we're interested in.
Dialogue: 0,0:14:21.52,0:14:24.20,Default,,0000,0000,0000,,So we do that for a while,\Nwe skip over a little bit,
Dialogue: 0,0:14:24.20,0:14:26.45,Default,,0000,0000,0000,,and then we train the\Nmachine learning algorithm
Dialogue: 0,0:14:26.45,0:14:28.42,Default,,0000,0000,0000,,based on these couple of hundred things,
Dialogue: 0,0:14:28.42,0:14:30.44,Default,,0000,0000,0000,,and we hope that it's gotten a lot better.
Dialogue: 0,0:14:30.44,0:14:33.52,Default,,0000,0000,0000,,You can see, it's now started to fade\Nsome of these pictures out,
Dialogue: 0,0:14:33.52,0:14:38.23,Default,,0000,0000,0000,,showing us that it already is recognizing\Nhow to understand some of these itself.
Dialogue: 0,0:14:38.23,0:14:41.13,Default,,0000,0000,0000,,We can then use this concept\Nof similar images,
Dialogue: 0,0:14:41.13,0:14:43.22,Default,,0000,0000,0000,,and using similar images, you can now see,
Dialogue: 0,0:14:43.22,0:14:47.24,Default,,0000,0000,0000,,the computer at this point is able to\Nentirely find just the fronts of cars.
Dialogue: 0,0:14:47.24,0:14:50.19,Default,,0000,0000,0000,,So at this point, the human\Ncan tell the computer,
Dialogue: 0,0:14:50.19,0:14:52.48,Default,,0000,0000,0000,,okay, yes, you've done\Na good job of that.
Dialogue: 0,0:14:53.65,0:14:55.84,Default,,0000,0000,0000,,Sometimes, of course, even at this point
Dialogue: 0,0:14:55.84,0:14:59.51,Default,,0000,0000,0000,,it's still difficult\Nto separate out groups.
Dialogue: 0,0:14:59.51,0:15:03.40,Default,,0000,0000,0000,,In this case, even after we let the\Ncomputer try to rotate this for a while,
Dialogue: 0,0:15:03.40,0:15:06.74,Default,,0000,0000,0000,,we still find that the left sides\Nand the right sides pictures
Dialogue: 0,0:15:06.74,0:15:08.22,Default,,0000,0000,0000,,are all mixed up together.
Dialogue: 0,0:15:08.22,0:15:10.36,Default,,0000,0000,0000,,So we can again give\Nthe computer some hints,
Dialogue: 0,0:15:10.36,0:15:13.34,Default,,0000,0000,0000,,and we say, okay, try and find\Na projection that separates out
Dialogue: 0,0:15:13.34,0:15:15.94,Default,,0000,0000,0000,,the left sides and the right sides\Nas much as possible
Dialogue: 0,0:15:15.94,0:15:18.07,Default,,0000,0000,0000,,using this deep learning algorithm.
Dialogue: 0,0:15:18.07,0:15:21.01,Default,,0000,0000,0000,,And giving it that hint --\Nah, okay, it's been successful.
Dialogue: 0,0:15:21.01,0:15:23.89,Default,,0000,0000,0000,,It's managed to find a way\Nof thinking about these objects
Dialogue: 0,0:15:23.89,0:15:26.27,Default,,0000,0000,0000,,that's separated out these together.
Dialogue: 0,0:15:26.27,0:15:28.71,Default,,0000,0000,0000,,So you get the idea here.
Dialogue: 0,0:15:28.71,0:15:36.91,Default,,0000,0000,0000,,This is a case not where the human\Nis being replaced by a computer,
Dialogue: 0,0:15:36.91,0:15:39.55,Default,,0000,0000,0000,,but where they're working together.
Dialogue: 0,0:15:39.55,0:15:43.10,Default,,0000,0000,0000,,What we're doing here is we're replacing\Nsomething that used to take a team
Dialogue: 0,0:15:43.10,0:15:45.10,Default,,0000,0000,0000,,of five or six people about seven years
Dialogue: 0,0:15:45.10,0:15:47.70,Default,,0000,0000,0000,,and replacing it with something\Nthat takes 15 minutes
Dialogue: 0,0:15:47.70,0:15:50.21,Default,,0000,0000,0000,,for one person acting alone.
Dialogue: 0,0:15:50.21,0:15:54.16,Default,,0000,0000,0000,,So this process takes about\Nfour or five iterations.
Dialogue: 0,0:15:54.16,0:15:56.02,Default,,0000,0000,0000,,You can see we now have 62 percent
Dialogue: 0,0:15:56.02,0:15:58.98,Default,,0000,0000,0000,,of our 1.5 million images \Nclassified correctly.
Dialogue: 0,0:15:58.98,0:16:01.45,Default,,0000,0000,0000,,And at this point, we\Ncan start to quite quickly
Dialogue: 0,0:16:01.45,0:16:02.74,Default,,0000,0000,0000,,grab whole big sections,
Dialogue: 0,0:16:02.74,0:16:05.66,Default,,0000,0000,0000,,check through them to make sure\Nthat there's no mistakes.
Dialogue: 0,0:16:05.66,0:16:09.62,Default,,0000,0000,0000,,Where there are mistakes, we can\Nlet the computer know about them.
Dialogue: 0,0:16:09.62,0:16:12.66,Default,,0000,0000,0000,,And using this kind of process\Nfor each of the different groups,
Dialogue: 0,0:16:12.66,0:16:15.15,Default,,0000,0000,0000,,we are now up to\Nan 80 percent success rate
Dialogue: 0,0:16:15.15,0:16:17.56,Default,,0000,0000,0000,,in classifying the 1.5 million images.
Dialogue: 0,0:16:17.56,0:16:19.64,Default,,0000,0000,0000,,And at this point, it's just a case
Dialogue: 0,0:16:19.64,0:16:23.22,Default,,0000,0000,0000,,of finding the small number\Nthat aren't classified correctly,
Dialogue: 0,0:16:23.22,0:16:26.11,Default,,0000,0000,0000,,and trying to understand why.
Dialogue: 0,0:16:26.11,0:16:27.85,Default,,0000,0000,0000,,And using that approach,
Dialogue: 0,0:16:27.85,0:16:31.97,Default,,0000,0000,0000,,by 15 minutes we get\Nto 97 percent classification rates.
Dialogue: 0,0:16:31.97,0:16:36.57,Default,,0000,0000,0000,,So this kind of technique\Ncould allow us to fix a major problem,
Dialogue: 0,0:16:36.58,0:16:39.61,Default,,0000,0000,0000,,which is that there's a lack\Nof medical expertise in the world.
Dialogue: 0,0:16:39.61,0:16:43.10,Default,,0000,0000,0000,,The World Economic Forum says\Nthat there's between a 10x and a 20x
Dialogue: 0,0:16:43.10,0:16:45.73,Default,,0000,0000,0000,,shortage of physicians\Nin the developing world,
Dialogue: 0,0:16:45.73,0:16:47.84,Default,,0000,0000,0000,,and it would take about 300 years
Dialogue: 0,0:16:47.84,0:16:50.73,Default,,0000,0000,0000,,to train enough people\Nto fix that problem.
Dialogue: 0,0:16:50.73,0:16:53.62,Default,,0000,0000,0000,,So imagine if we can help\Nenhance their efficiency
Dialogue: 0,0:16:53.62,0:16:56.46,Default,,0000,0000,0000,,using these deep learning approaches?
Dialogue: 0,0:16:56.46,0:16:58.69,Default,,0000,0000,0000,,So I'm very excited\Nabout the opportunities.
Dialogue: 0,0:16:58.69,0:17:01.28,Default,,0000,0000,0000,,I'm also concerned about the problems.
Dialogue: 0,0:17:01.28,0:17:04.40,Default,,0000,0000,0000,,The problem here is that\Nevery area in blue on this map
Dialogue: 0,0:17:04.40,0:17:08.17,Default,,0000,0000,0000,,is somewhere where services\Nare over 80 percent of employment.
Dialogue: 0,0:17:08.17,0:17:09.96,Default,,0000,0000,0000,,What are services?
Dialogue: 0,0:17:09.96,0:17:11.47,Default,,0000,0000,0000,,These are services.
Dialogue: 0,0:17:11.47,0:17:15.63,Default,,0000,0000,0000,,These are also the exact things that\Ncomputers have just learned how to do.
Dialogue: 0,0:17:15.63,0:17:19.43,Default,,0000,0000,0000,,So 80 percent of the world's employment\Nin the developed world
Dialogue: 0,0:17:19.43,0:17:21.96,Default,,0000,0000,0000,,is stuff that computers \Nhave just learned how to do.
Dialogue: 0,0:17:21.96,0:17:23.40,Default,,0000,0000,0000,,What does that mean?
Dialogue: 0,0:17:23.40,0:17:25.99,Default,,0000,0000,0000,,Well, it'll be fine.\NThey'll be replaced by other jobs.
Dialogue: 0,0:17:25.99,0:17:28.69,Default,,0000,0000,0000,,For example, there will be\Nmore jobs for data scientists.
Dialogue: 0,0:17:28.69,0:17:29.51,Default,,0000,0000,0000,,Well, not really.
Dialogue: 0,0:17:29.51,0:17:32.63,Default,,0000,0000,0000,,It doesn't take data scientists \Nvery long to build these things.
Dialogue: 0,0:17:32.63,0:17:35.88,Default,,0000,0000,0000,,For example, these four algorithms\Nwere all built by the same guy.
Dialogue: 0,0:17:35.88,0:17:38.32,Default,,0000,0000,0000,,So if you think, oh, \Nit's all happened before,
Dialogue: 0,0:17:38.32,0:17:42.13,Default,,0000,0000,0000,,we've seen the results in the past\Nof when new things come along
Dialogue: 0,0:17:42.13,0:17:44.38,Default,,0000,0000,0000,,and they get replaced by new jobs,
Dialogue: 0,0:17:44.38,0:17:46.49,Default,,0000,0000,0000,,what are these new jobs going to be?
Dialogue: 0,0:17:46.49,0:17:48.36,Default,,0000,0000,0000,,It's very hard for us to estimate this,
Dialogue: 0,0:17:48.36,0:17:51.10,Default,,0000,0000,0000,,because human performance\Ngrows at this gradual rate,
Dialogue: 0,0:17:51.10,0:17:53.67,Default,,0000,0000,0000,,but we now have a system, deep learning,
Dialogue: 0,0:17:53.67,0:17:56.89,Default,,0000,0000,0000,,that we know actually grows\Nin capability exponentially.
Dialogue: 0,0:17:56.89,0:17:58.50,Default,,0000,0000,0000,,And we're here.
Dialogue: 0,0:17:58.50,0:18:00.56,Default,,0000,0000,0000,,So currently, we see the things around us
Dialogue: 0,0:18:00.56,0:18:03.24,Default,,0000,0000,0000,,and we say, "Oh, computers\Nare still pretty dumb." Right?
Dialogue: 0,0:18:03.24,0:18:06.66,Default,,0000,0000,0000,,But in five years' time,\Ncomputers will be off this chart.
Dialogue: 0,0:18:06.66,0:18:10.53,Default,,0000,0000,0000,,So we need to be starting to think\Nabout this capability right now.
Dialogue: 0,0:18:10.53,0:18:12.58,Default,,0000,0000,0000,,We have seen this once before, of course.
Dialogue: 0,0:18:12.58,0:18:13.97,Default,,0000,0000,0000,,In the Industrial Revolution,
Dialogue: 0,0:18:13.97,0:18:16.82,Default,,0000,0000,0000,,we saw a step change\Nin capability thanks to engines.
Dialogue: 0,0:18:17.67,0:18:20.80,Default,,0000,0000,0000,,The thing is, though,\Nthat after a while, things flattened out.
Dialogue: 0,0:18:20.80,0:18:22.51,Default,,0000,0000,0000,,There was social disruption,
Dialogue: 0,0:18:22.51,0:18:25.95,Default,,0000,0000,0000,,but once engines were used \Nto generate power in all the situations,
Dialogue: 0,0:18:25.95,0:18:28.30,Default,,0000,0000,0000,,things really settled down.
Dialogue: 0,0:18:28.30,0:18:29.77,Default,,0000,0000,0000,,The Machine Learning Revolution
Dialogue: 0,0:18:29.77,0:18:32.68,Default,,0000,0000,0000,,is going to be very different\Nfrom the Industrial Revolution,
Dialogue: 0,0:18:32.68,0:18:35.63,Default,,0000,0000,0000,,because the Machine Learning Revolution,\Nit never settles down.
Dialogue: 0,0:18:35.63,0:18:38.61,Default,,0000,0000,0000,,The better computers get\Nat intellectual activities,
Dialogue: 0,0:18:38.61,0:18:42.86,Default,,0000,0000,0000,,the more they can build better computers\Nto be better at intellectual capabilities,
Dialogue: 0,0:18:42.86,0:18:44.77,Default,,0000,0000,0000,,so this is going to be a kind of change
Dialogue: 0,0:18:44.77,0:18:47.25,Default,,0000,0000,0000,,that the world has actually\Nnever experienced before,
Dialogue: 0,0:18:47.25,0:18:50.55,Default,,0000,0000,0000,,so your previous understanding\Nof what's possible is different.
Dialogue: 0,0:18:50.97,0:18:52.75,Default,,0000,0000,0000,,This is already impacting us.
Dialogue: 0,0:18:52.75,0:18:56.38,Default,,0000,0000,0000,,In the last 25 years,\Nas capital productivity has increased,
Dialogue: 0,0:18:56.40,0:19:00.59,Default,,0000,0000,0000,,labor productivity has been flat,\Nin fact even a little bit down.
Dialogue: 0,0:19:01.41,0:19:04.15,Default,,0000,0000,0000,,So I want us to start\Nhaving this discussion now.
Dialogue: 0,0:19:04.15,0:19:07.18,Default,,0000,0000,0000,,I know that when I often tell people\Nabout this situation,
Dialogue: 0,0:19:07.18,0:19:08.67,Default,,0000,0000,0000,,people can be quite dismissive.
Dialogue: 0,0:19:08.67,0:19:10.34,Default,,0000,0000,0000,,Well, computers can't really think,
Dialogue: 0,0:19:10.34,0:19:13.37,Default,,0000,0000,0000,,they don't emote,\Nthey don't understand poetry,
Dialogue: 0,0:19:13.37,0:19:15.89,Default,,0000,0000,0000,,we don't really understand how they work.
Dialogue: 0,0:19:15.89,0:19:17.37,Default,,0000,0000,0000,,So what?
Dialogue: 0,0:19:17.37,0:19:19.18,Default,,0000,0000,0000,,Computers right now can do the things
Dialogue: 0,0:19:19.18,0:19:21.90,Default,,0000,0000,0000,,that humans spend most\Nof their time being paid to do,
Dialogue: 0,0:19:21.90,0:19:23.63,Default,,0000,0000,0000,,so now's the time to start thinking
Dialogue: 0,0:19:23.63,0:19:28.02,Default,,0000,0000,0000,,about how we're going to adjust our\Nsocial structures and economic structures
Dialogue: 0,0:19:28.02,0:19:29.86,Default,,0000,0000,0000,,to be aware of this new reality.
Dialogue: 0,0:19:29.86,0:19:31.39,Default,,0000,0000,0000,,Thank you.
Dialogue: 0,0:19:31.39,0:19:32.19,Default,,0000,0000,0000,,(Applause)