YouTube

Got a YouTube account?

New: enable viewer-created translations and captions on your YouTube channel!

English subtitles

← The wonderful and terrifying implications of computers that can learn

Get Embed Code
30 Languages

Showing Revision 8 created 12/16/2014 by Morton Bast.

  1. It used to be that if you wanted
    to get a computer to do something new,
  2. you would have to program it.
  3. Now, programming, for those of you here
    that haven't done it yourself,
  4. requires laying out in excruciating detail
  5. every single step that you want
    the computer to do
  6. in order to achieve your goal.
  7. Now, if you want to do something
    that you don't know how to do yourself,
  8. then this is going
    to be a great challenge.
  9. So this was the challenge faced
    by this man, Arthur Samuel.

  10. In 1956, he wanted to get this computer
  11. to be able to beat him at checkers.
  12. How can you write a program,
  13. lay out in excruciating detail,
    how to be better than you at checkers?
  14. So he came up with an idea:
  15. he had the computer play
    against itself thousands of times
  16. and learn how to play checkers.
  17. And indeed it worked,
    and in fact, by 1962,
  18. this computer had beaten
    the Connecticut state champion.
  19. So Arthur Samuel was
    the father of machine learning,

  20. and I have a great debt to him,
  21. because I am a machine
    learning practitioner.
  22. I was the president of Kaggle,
  23. a community of over 200,000
    machine learning practictioners.
  24. Kaggle puts up competitions
  25. to try and get them to solve
    previously unsolved problems,
  26. and it's been successful
    hundreds of times.
  27. So from this vantage point,
    I was able to find out
  28. a lot about what machine learning
    can do in the past, can do today,
  29. and what it could do in the future.
  30. Perhaps the first big success of
    machine learning commercially was Google.
  31. Google showed that it is
    possible to find information
  32. by using a computer algorithm,
  33. and this algorithm is based
    on machine learning.
  34. Since that time, there have been many
    commercial successes of machine learning.
  35. Companies like Amazon and Netflix
  36. use machine learning to suggest
    products that you might like to buy,
  37. movies that you might like to watch.
  38. Sometimes, it's almost creepy.
  39. Companies like LinkedIn and Facebook
  40. sometimes will tell you about
    who your friends might be
  41. and you have no idea how it did it,
  42. and this is because it's using
    the power of machine learning.
  43. These are algorithms that have
    learned how to do this from data
  44. rather than being programmed by hand.
  45. This is also how IBM was successful

  46. in getting Watson to beat
    the two world champions at "Jeopardy,"
  47. answering incredibly subtle
    and complex questions like this one.
  48. ["The ancient 'Lion of Nimrud' went missing
    from this city's national museum in 2003
    (along with a lot of other stuff)"]
  49. This is also why we are now able
    to see the first self-driving cars.
  50. If you want to be able to tell
    the difference between, say,
  51. a tree and a pedestrian,
    well, that's pretty important.
  52. We don't know how to write
    those programs by hand,
  53. but with machine learning,
    this is now possible.
  54. And in fact, this car has driven
    over a million miles
  55. without any accidents on regular roads.
  56. So we now know that computers can learn,

  57. and computers can learn to do things
  58. that we actually sometimes
    don't know how to do ourselves,
  59. or maybe can do them better than us.
  60. One of the most amazing examples
    I've seen of machine learning
  61. happened on a project that I ran at Kaggle
  62. where a team run by a guy
    called Geoffrey Hinton
  63. from the University of Toronto
  64. won a competition for
    automatic drug discovery.
  65. Now, what was extraordinary here
    is not just that they beat
  66. all of the algorithms developed by Merck
    or the international academic community,
  67. but nobody on the team had any background
    in chemistry or biology or life sciences,
  68. and they did it in two weeks.
  69. How did they do this?
  70. They used an extraordinary algorithm
    called deep learning.
  71. So important was this that in fact
    the success was covered
  72. in The New York Times in a front page
    article a few weeks later.
  73. This is Geoffrey Hinton
    here on the left-hand side.
  74. Deep learning is an algorithm
    inspired by how the human brain works,
  75. and as a result it's an algorithm
  76. which has no theoretical limitations
    on what it can do.
  77. The more data you give it and the more
    computation time you give it,
  78. the better it gets.
  79. The New York Times also
    showed in this article

  80. another extraordinary
    result of deep learning
  81. which I'm going to show you now.
  82. It shows that computers
    can listen and understand.
  83. (Video) Richard Rashid: Now, the last step

  84. that I want to be able
    to take in this process
  85. is to actually speak to you in Chinese.
  86. Now the key thing there is,
  87. we've been able to take a large amount
    of information from many Chinese speakers
  88. and produce a text-to-speech system
  89. that takes Chinese text
    and converts it into Chinese language,
  90. and then we've taken
    an hour or so of my own voice
  91. and we've used that to modulate
  92. the standard text-to-speech system
    so that it would sound like me.
  93. Again, the result's not perfect.
  94. There are in fact quite a few errors.
  95. (In Chinese)
  96. (Applause)
  97. There's much work to be done in this area.
  98. (In Chinese)
  99. (Applause)
  100. Jeremy Howard: Well, that was at
    a machine learning conference in China.

  101. It's not often, actually,
    at academic conferences
  102. that you do hear spontaneous applause,
  103. although of course sometimes
    at TEDx conferences, feel free.
  104. Everything you saw there
    was happening with deep learning.
  105. (Applause) Thank you.
  106. The transcription in English
    was deep learning.
  107. The translation to Chinese and the text
    in the top right, deep learning,
  108. and the construction of the voice
    was deep learning as well.
  109. So deep learning is
    this extraordinary thing.

  110. It's a single algorithm that
    can seem to do almost anything,
  111. and I discovered that a year earlier,
    it had also learned to see.
  112. In this obscure competition from Germany
  113. called the German Traffic Sign
    Recognition Benchmark,
  114. deep learning had learned
    to recognize traffic signs like this one.
  115. Not only could it
    recognize the traffic signs
  116. better than any other algorithm,
  117. the leaderboard actually showed
    it was better than people,
  118. about twice as good as people.
  119. So by 2011, we had the first example
  120. of computers that can see
    better than people.
  121. Since that time, a lot has happened.
  122. In 2012, Google announced that
    they had a deep learning algorithm
  123. watch YouTube videos
  124. and crunched the data
    on 16,000 computers for a month,
  125. and the computer independently learned
    about concepts such as people and cats
  126. just by watching the videos.
  127. This is much like the way
    that humans learn.
  128. Humans don't learn
    by being told what they see,
  129. but by learning for themselves
    what these things are.
  130. Also in 2012, Geoffrey Hinton,
    who we saw earlier,
  131. won the very popular ImageNet competition,
  132. looking to try to figure out
    from one and a half million images
  133. what they're pictures of.
  134. As of 2014, we're now down
    to a six percent error rate
  135. in image recognition.
  136. This is better than people, again.
  137. So machines really are doing
    an extraordinarily good job of this,

  138. and it is now being used in industry.
  139. For example, Google announced last year
  140. that they had mapped every single
    location in France in two hours,
  141. and the way they did it was
    that they fed street view images
  142. into a deep learning algorithm
    to recognize and read street numbers.
  143. Imagine how long
    it would have taken before:
  144. dozens of people, many years.
  145. This is also happening in China.
  146. Baidu is kind of
    the Chinese Google, I guess,
  147. and what you see here in the top left
  148. is an example of a picture that I uploaded
    to Baidu's deep learning system,
  149. and underneath you can see that the system
    has understood what that picture is
  150. and found similar images.
  151. The similar images actually
    have similar backgrounds,
  152. similar directions of the faces,
  153. even some with their tongue out.
  154. This is not clearly looking
    at the text of a web page.
  155. All I uploaded was an image.
  156. So we now have computers which
    really understand what they see
  157. and can therefore search databases
  158. of hundreds of millions
    of images in real time.
  159. So what does it mean
    now that computers can see?

  160. Well, it's not just
    that computers can see.
  161. In fact, deep learning
    has done more than that.
  162. Complex, nuanced sentences like this one
  163. are now understandable
    with deep learning algorithms.
  164. As you can see here,
  165. this Stanford-based system
    showing the red dot at the top
  166. has figured out that this sentence
    is expressing negative sentiment.
  167. Deep learning now in fact
    is near human performance
  168. at understanding what sentences are about
    and what it is saying about those things.
  169. Also, deep learning has
    been used to read Chinese,
  170. again at about native
    Chinese speaker level.
  171. This algorithm developed
    out of Switzerland
  172. by people, none of whom speak
    or understand any Chinese.
  173. As I say, using deep learning
  174. is about the best system
    in the world for this,
  175. even compared to native
    human understanding.
  176. This is a system that we
    put together at my company

  177. which shows putting
    all this stuff together.
  178. These are pictures which
    have no text attached,
  179. and as I'm typing in here sentences,
  180. in real time it's understanding
    these pictures
  181. and figuring out what they're about
  182. and finding pictures that are similar
    to the text that I'm writing.
  183. So you can see, it's actually
    understanding my sentences
  184. and actually understanding these pictures.
  185. I know that you've seen
    something like this on Google,
  186. where you can type in things
    and it will show you pictures,
  187. but actually what it's doing is it's
    searching the webpage for the text.
  188. This is very different from actually
    understanding the images.
  189. This is something that computers
    have only been able to do
  190. for the first time in the last few months.
  191. So we can see now that computers
    can not only see but they can also read,

  192. and, of course, we've shown that they
    can understand what they hear.
  193. Perhaps not surprising now that
    I'm going to tell you they can write.
  194. Here is some text that I generated
    using a deep learning algorithm yesterday.
  195. And here is some text that an algorithm
    out of Stanford generated.
  196. Each of these sentences was generated
  197. by a deep learning algorithm
    to describe each of those pictures.
  198. This algorithm before has never seen
    a man in a black shirt playing a guitar.
  199. It's seen a man before,
    it's seen black before,
  200. it's seen a guitar before,
  201. but it has independently generated
    this novel description of this picture.
  202. We're still not quite at human
    performance here, but we're close.
  203. In tests, humans prefer
    the computer-generated caption
  204. one out of four times.
  205. Now this system is now only two weeks old,
  206. so probably within the next year,
  207. the computer algorithm will be
    well past human performance
  208. at the rate things are going.
  209. So computers can also write.
  210. So we put all this together and it leads
    to very exciting opportunities.

  211. For example, in medicine,
  212. a team in Boston announced
    that they had discovered
  213. dozens of new clinically relevant features
  214. of tumors which help doctors
    make a prognosis of a cancer.
  215. Very similarly, in Stanford,
  216. a group there announced that,
    looking at tissues under magnification,
  217. they've developed
    a machine learning-based system
  218. which in fact is better
    than human pathologists
  219. at predicting survival rates
    for cancer sufferers.
  220. In both of these cases, not only
    were the predictions more accurate,
  221. but they generated new insightful science.
  222. In the radiology case,
  223. they were new clinical indicators
    that humans can understand.
  224. In this pathology case,
  225. the computer system actually discovered
    that the cells around the cancer
  226. are as important as
    the cancer cells themselves
  227. in making a diagnosis.
  228. This is the opposite of what pathologists
    had been taught for decades.
  229. In each of those two cases,
    they were systems developed
  230. by a combination of medical experts
    and machine learning experts,
  231. but as of last year,
    we're now beyond that too.
  232. This is an example of
    identifying cancerous areas
  233. of human tissue under a microscope.
  234. The system being shown here
    can identify those areas more accurately,
  235. or about as accurately,
    as human pathologists,
  236. but was built entirely with deep learning
    using no medical expertise
  237. by people who have
    no background in the field.
  238. Similarly, here, this neuron segmentation.
  239. We can now segment neurons
    about as accurately as humans can,
  240. but this system was developed
    with deep learning
  241. using people with no previous
    background in medicine.
  242. So myself, as somebody with
    no previous background in medicine,

  243. I seem to be entirely well qualified
    to start a new medical company,
  244. which I did.
  245. I was kind of terrified of doing it,
  246. but the theory seemed to suggest
    that it ought to be possible
  247. to do very useful medicine
    using just these data analytic techniques.
  248. And thankfully, the feedback
    has been fantastic,
  249. not just from the media
    but from the medical community,
  250. who have been very supportive.
  251. The theory is that we can take
    the middle part of the medical process
  252. and turn that into data analysis
    as much as possible,
  253. leaving doctors to do
    what they're best at.
  254. I want to give you an example.
  255. It now takes us about 15 minutes
    to generate a new medical diagnostic test
  256. and I'll show you that in real time now,
  257. but I've compressed it down to
    three minutes by cutting some pieces out.
  258. Rather than showing you
    creating a medical diagnostic test,
  259. I'm going to show you
    a diagnostic test of car images,
  260. because that's something
    we can all understand.
  261. So here we're starting with
    about 1.5 million car images,

  262. and I want to create something
    that can split them into the angle
  263. of the photo that's being taken.
  264. So these images are entirely unlabeled,
    so I have to start from scratch.
  265. With our deep learning algorithm,
  266. it can automatically identify
    areas of structure in these images.
  267. So the nice thing is that the human
    and the computer can now work together.
  268. So the human, as you can see here,
  269. is telling the computer
    about areas of interest
  270. which it wants the computer then
    to try and use to improve its algorithm.
  271. Now, these deep learning systems actually
    are in 16,000-dimensional space,
  272. so you can see here the computer
    rotating this through that space,
  273. trying to find new areas of structure.
  274. And when it does so successfully,
  275. the human who is driving it can then
    point out the areas that are interesting.
  276. So here, the computer has
    successfully found areas,
  277. for example, angles.
  278. So as we go through this process,
  279. we're gradually telling
    the computer more and more
  280. about the kinds of structures
    we're looking for.
  281. You can imagine in a diagnostic test
  282. this would be a pathologist identifying
    areas of pathosis, for example,
  283. or a radiologist indicating
    potentially troublesome nodules.
  284. And sometimes it can be
    difficult for the algorithm.
  285. In this case, it got kind of confused.
  286. The fronts and the backs
    of the cars are all mixed up.
  287. So here we have to be a bit more careful,
  288. manually selecting these fronts
    as opposed to the backs,
  289. then telling the computer
    that this is a type of group
  290. that we're interested in.
  291. So we do that for a while,
    we skip over a little bit,

  292. and then we train the
    machine learning algorithm
  293. based on these couple of hundred things,
  294. and we hope that it's gotten a lot better.
  295. You can see, it's now started to fade
    some of these pictures out,
  296. showing us that it already is recognizing
    how to understand some of these itself.
  297. We can then use this concept
    of similar images,
  298. and using similar images, you can now see,
  299. the computer at this point is able to
    entirely find just the fronts of cars.
  300. So at this point, the human
    can tell the computer,
  301. okay, yes, you've done
    a good job of that.
  302. Sometimes, of course, even at this point

  303. it's still difficult
    to separate out groups.
  304. In this case, even after we let the
    computer try to rotate this for a while,
  305. we still find that the left sides
    and the right sides pictures
  306. are all mixed up together.
  307. So we can again give
    the computer some hints,
  308. and we say, okay, try and find
    a projection that separates out
  309. the left sides and the right sides
    as much as possible
  310. using this deep learning algorithm.
  311. And giving it that hint --
    ah, okay, it's been successful.
  312. It's managed to find a way
    of thinking about these objects
  313. that's separated out these together.
  314. So you get the idea here.

  315. This is a case not where the human
    is being replaced by a computer,
  316. but where they're working together.
  317. What we're doing here is we're replacing
    something that used to take a team
  318. of five or six people about seven years
  319. and replacing it with something
    that takes 15 minutes
  320. for one person acting alone.
  321. So this process takes about
    four or five iterations.

  322. You can see we now have 62 percent
  323. of our 1.5 million images
    classified correctly.
  324. And at this point, we
    can start to quite quickly
  325. grab whole big sections,
  326. check through them to make sure
    that there's no mistakes.
  327. Where there are mistakes, we can
    let the computer know about them.
  328. And using this kind of process
    for each of the different groups,
  329. we are now up to
    an 80 percent success rate
  330. in classifying the 1.5 million images.
  331. And at this point, it's just a case
  332. of finding the small number
    that aren't classified correctly,
  333. and trying to understand why.
  334. And using that approach,
  335. by 15 minutes we get
    to 97 percent classification rates.
  336. So this kind of technique
    could allow us to fix a major problem,

  337. which is that there's a lack
    of medical expertise in the world.
  338. The World Economic Forum says
    that there's between a 10x and a 20x
  339. shortage of physicians
    in the developing world,
  340. and it would take about 300 years
  341. to train enough people
    to fix that problem.
  342. So imagine if we can help
    enhance their efficiency
  343. using these deep learning approaches?
  344. So I'm very excited
    about the opportunities.

  345. I'm also concerned about the problems.
  346. The problem here is that
    every area in blue on this map
  347. is somewhere where services
    are over 80 percent of employment.
  348. What are services?
  349. These are services.
  350. These are also the exact things that
    computers have just learned how to do.
  351. So 80 percent of the world's employment
    in the developed world
  352. is stuff that computers
    have just learned how to do.
  353. What does that mean?
  354. Well, it'll be fine.
    They'll be replaced by other jobs.
  355. For example, there will be
    more jobs for data scientists.
  356. Well, not really.
  357. It doesn't take data scientists
    very long to build these things.
  358. For example, these four algorithms
    were all built by the same guy.
  359. So if you think, oh,
    it's all happened before,
  360. we've seen the results in the past
    of when new things come along
  361. and they get replaced by new jobs,
  362. what are these new jobs going to be?
  363. It's very hard for us to estimate this,
  364. because human performance
    grows at this gradual rate,
  365. but we now have a system, deep learning,
  366. that we know actually grows
    in capability exponentially.
  367. And we're here.
  368. So currently, we see the things around us
  369. and we say, "Oh, computers
    are still pretty dumb." Right?
  370. But in five years' time,
    computers will be off this chart.
  371. So we need to be starting to think
    about this capability right now.
  372. We have seen this once before, of course.

  373. In the Industrial Revolution,
  374. we saw a step change
    in capability thanks to engines.
  375. The thing is, though,
    that after a while, things flattened out.
  376. There was social disruption,
  377. but once engines were used
    to generate power in all the situations,
  378. things really settled down.
  379. The Machine Learning Revolution
  380. is going to be very different
    from the Industrial Revolution,
  381. because the Machine Learning Revolution,
    it never settles down.
  382. The better computers get
    at intellectual activities,
  383. the more they can build better computers
    to be better at intellectual capabilities,
  384. so this is going to be a kind of change
  385. that the world has actually
    never experienced before,
  386. so your previous understanding
    of what's possible is different.
  387. This is already impacting us.

  388. In the last 25 years,
    as capital productivity has increased,
  389. labor productivity has been flat,
    in fact even a little bit down.
  390. So I want us to start
    having this discussion now.

  391. I know that when I often tell people
    about this situation,
  392. people can be quite dismissive.
  393. Well, computers can't really think,
  394. they don't emote,
    they don't understand poetry,
  395. we don't really understand how they work.
  396. So what?
  397. Computers right now can do the things
  398. that humans spend most
    of their time being paid to do,
  399. so now's the time to start thinking
  400. about how we're going to adjust our
    social structures and economic structures
  401. to be aware of this new reality.
  402. Thank you.
  403. (Applause)