English subtitles

← True Artificial Intelligence will change everything | Juergen Schmidhuber | TEDxLakeComo

Get Embed Code
7 Languages

Showing Revision 10 created 08/24/2019 by Saeed Hosseinzadeh.

  1. When I was a boy,
  2. I wanted to maximise
    my impact on the world,
  3. and I was smart enough
    to realise that I am not very smart.
  4. And that I have to build a machine
  5. that learns to become
    much smarter than myself,
  6. such that it can solve all the problems
    that I cannot solve myself,
  7. and I can retire.
  8. And my first publication
    on that dates back 30 years: 1987.
  9. My diploma thesis,
  10. where I already try to solve
    the grand problem of AI,
  11. not only build a machine
  12. that learns a little bit here,
    learns a little bit there,
  13. but also learns to improve
    the learning algorithm itself.
  14. And the way it learns, the way it learns,
  15. and so on recursively, without any limits
  16. except the limits of logics and physics.
  17. And, I'm still working
    on the same old thing,
  18. and I'm still pretty much
    saying the same thing,
  19. except that now
    more people are listening.
  20. Because the learning algorithms
  21. that we have developed
    on the way to this goal,
  22. they are now on 3.000 million smartphones.
  23. And all of you have them in your pockets.
  24. What you see here
  25. are the five most valuable companies
    of the Western world:
  26. Apple, Google, Facebook,
    Microsoft and Amazon.
  27. And all of them are emphasising
  28. that AI, artificial intelligence,
  29. is central to what they are doing.
  30. And all of them are using heavily
    the deep learning methods
  31. that my team has developed
    since the early nineties,
  32. in Munich and in Switzerland.
  33. Especially something which is called:
    "the long short-term memory".
  34. Has anybody in this room ever heard
    of the long short-term memory,
  35. or the LSTM?
  36. Hands up, anybody ever heard of that?
  37. Okay.
  38. Has anybody never heard of the LSTM?
  39. Okay.
    I see we have a third group in this room:
  40. [those] who didn't
    understand the question.
  41. (Laughter)
  42. The LSTM is a little bit like your brain:
  43. it's an artificial neural network
    which also has neurons,
  44. and in your brain, you've got
    about 100 billion neurons.
  45. And each of them is connected
  46. to roughly 10,000
    other neurons on average,
  47. Which means that you have got
    a million billion connections.
  48. And each of these connections
    has a "strength"
  49. which says how much
    does this neuron over here
  50. influence that one over there
    at the next time step.
  51. And in the beginning,
  52. all these connections are random
    and the system knows nothing;
  53. but then, through a smart
    learning algorithm,
  54. it learns from lots of examples
    to translate the incoming data,
  55. such as video through the cameras,
    or audio through the microphones,
  56. or pain signals through the pain sensors.
  57. It learns to translate that
    into output actions,
  58. because some of these neurons
    are output neurons,
  59. that control speech muscles
    and finger muscles.
  60. And only through experience,
  61. it can learn to solve
    all kinds of interesting problems,
  62. such as driving a car
  63. or do the speech recognition
    on your smartphone.
  64. Because whenever you take out
    your smartphone,
  65. an Android phone, for example,
  66. and you speak to it, and you say:
  67. "Ok Google, show me
    the shortest way to Milano."
  68. Then it understands your speech.
  69. Because there is a LSTM in there
    which has learned to understand speech.
  70. Every ten milliseconds,
    100 times a second,
  71. new inputs are coming from the microphone,
  72. and then are translated, after thinking,
  73. into letters
  74. which are then questioned
    to the search engine.
  75. And it has learned to do that
  76. by listening to lots of speech
    from women, from men, all kinds of people.
  77. And that's how, since 2015,
  78. Google speech recognition
    is now much better than it used to be.
  79. The basic LSTM cell looks like that:
  80. I don't have the time to explain that,
  81. but at least I can list the names
  82. of the brilliant students in my lab
    who made that possible.
  83. And what are the big companies
    doing with that?
  84. Well, speech recognition
    is only one example;
  85. if you are on Facebook -
    is anybody on Facebook?
  86. Are you sometimes clicking
    at the translate button?
  87. because somebody sent you something
    in a foreign language
  88. and then you can translate it.
  89. Is anybody doing that? Yeah.
  90. Whenever you do that,
  91. you are waking up, again,
    a long short term memory, an LSTM,
  92. which has learned to translate
    text in one language
  93. into translated text.
  94. And Facebook is doing that
    four billion times a day,
  95. so every second 50,000 sentences
  96. are being translated
  97. by an LSTM working for Facebook;
  98. and another 50,000 in the second;
    then another 50,000.
  99. And to see how much this thing
    is now permitting the modern world,
  100. just note that almost 30 percent
  101. of the awesome computational
    power for inference
  102. and all these Google Data Centers,
  103. all these data centers of Google,
    all over the world,
  104. is used for LSTM.
  105. Almost 30 percent.
  106. If you have an Amazon Echo,
  107. you can ask a question and it answers you.
  108. And the voice that you hear
    it's not a recording;
  109. it's an LSTM network
  110. which has learned from training examples
  111. to sound like a female voice.
  112. If you have an iPhone,
    and you're using the quick type,
  113. it's trying to predict
    what you want to do next
  114. given all the previous context
    of what you did so far.
  115. Again, that's an LSTM
    which has learned to do that,
  116. so it's on a billion iPhones.
  117. You are a large audience, by my standards:
  118. but when we started this work,
    decades ago, in the early '90s,
  119. only few people were interested in that,
  120. because computers were so slow
    and you couldn't do so much with it.
  121. And I remember I gave a talk
    at a conference,
  122. and there was just
    one single person in the audience,
  123. a young lady.
  124. I said, young lady,
    it's very embarrassing,
  125. but apparently today
    I'm going to give this talk just to you.
  126. And she said,
  127. "OK, but please hurry:
    I am the next speaker!"
  128. (Laughter)
  129. Since then, we have
    greatly profited from the fact
  130. that every five years
    computers are getting ten times cheaper,
  131. which is an old trend that has held
    since 1941 at least.
  132. Since this man, Konrad Zuse,
  133. built the first working
    program controlled computer in Berlin
  134. and he could do, roughly,
    one operation per second.
  135. One!
  136. And then ten years later,
    for the same price,
  137. one could do 100 operations:
  138. 30 years later,
  139. 1 million operations for the same price;
  140. and today, after 75 years, we can do
  141. a million billion times as much
    for the same price.
  142. And the trend is not about to stop,
  143. because the physical limits
    are much further out there.
  144. Rather soon, and not
    so many years or decades,
  145. we will for the first time
    have little computational devices
  146. that can compute as much as a human brain;
  147. and that's a trend that doesn't break.
  148. 50 years later, there will be
    a little computational device,
  149. for the same price,
  150. that can compute as much as all
    10 billion human brains taken together.
  151. and there will not only be one,
    of those devices, but many many many.
  152. Everything is going to change.
  153. Already in 2011,
    computers were fast enough
  154. such that our deep learning methods
  155. for the first time could achieve
    a superhuman pattern-recognition result.
  156. It was the first superhuman result
    in the history of computer vision.
  157. And back then, computers were
    20 times more expensive than today.
  158. So today, for the same price,
  159. we can do 20 times as much.
  160. And just five years ago,
  161. when computers were 10 times
    more expensive than today,
  162. we already could win, for the first time,
    medical imaging competitions.
  163. What you see behind me
    is a slice through the female breast
  164. and the tissue that you see there
    has all kinds of cells;
  165. and normally you need a trained doctor,
    a trained histologist
  166. who is able to detect
    the dangerous cancer cells,
  167. or pre-cancer cells.
  168. Now, our stupid network
  169. knows nothing about cancer,
    knows nothing about vision.
  170. It knows nothing in the beginning:
  171. but we can train it to imitate
    the human teacher, the doctor.
  172. And it became as good, or better,
    than the best competitors.
  173. And very soon,
  174. all of medical diagnosis
    is going to be superhuman.
  175. And it's going to be mandatory,
  176. because it's going to be
    so much better than the doctors.
  177. After this, all kinds of medical
    imaging startups were founded
  178. focusing just on this,
    because it's so important.
  179. We can also use LSTM to train robots.
  180. One important thing I want to say is,
  181. that we not only have systems
  182. that slavishly imitate
    what humans show them;
  183. no, we also have AIs
    that set themselves their own goals.
  184. And like little babies,
    invent their own experiment
  185. to explore the world
  186. and to figure out
    what you can do in the world.
  187. Without a teacher.
  188. And becoming more and more general
    problem solvers in the process,
  189. by learning new skills
    on top of old skills.
  190. And this is going to scale:
    we call that "Artificial Curiosity".
  191. Or a recent buzzword is "power plane".
  192. Learning to become a more and more
    general problem solvers
  193. by learning to invent, like a scientist,
    one new interesting goal after another.
  194. And it's going to scale.
  195. And I think,
  196. in not so many years
    from now, for the first time,
  197. we are going to have an animal-like AI -
  198. we don't have that yet.
  199. On the level of a little crow,
  200. which already can learn
    to use tools, for example,
  201. or a little monkey.
  202. And once we have that,
  203. it may take just a few decades
  204. to do the final step
    towards human level intelligence.
  205. Because technological evolution
  206. is about a million times faster
    than biological evolution,
  207. and biological evolution
    needed 3.5 billion years
  208. to evolve a monkey from scratch.
  209. But then, it took just a few tens
    of millions of years afterwards
  210. to evolve human level intelligence.
  211. We have a company
    which is called Nnaisense
  212. like birth in [French], "Naissance",
    but spelled in a different way,
  213. which is trying to make this a reality
  214. and build the first
    true general-purpose AI.
  215. At the moment, almost all research in AI
    is very human centric,
  216. and it's all about making human lives
    longer and healthier and easier
  217. and making humans
    more addicted to their smartphones.
  218. But in the long run, AIs are going to -
    especially the smart ones -
  219. are going to set themselves
    their own goals.
  220. And I have no doubt, in my mind,
  221. that they are going to become
    much smarter than we are.
  222. And what are they going to do?
  223. Of course they are going to realize
    what we have realized a long time ago;
  224. namely, that most of the resources,
    in the solar system or in general,
  225. are not in our little biosphere.
  226. They are out there in space.
  227. And so, of course,
    they are going to emigrate.
  228. And of course they are going to use
  229. trillions of self-replicating
    robot factories
  230. to expand in form of a growing AI bubble
  231. which within a few hundred thousand years
  232. is going to cover the entire galaxy
  233. by senders and receivers
  234. such that AIs can travel
  235. the way they are
    already traveling in my lab:
  236. by radio, from sender to receiver.
  237. Wireless.
  238. So what we are witnessing now
  239. is much more than just
    another Industrial Revolution.
  240. This is something
    that transcends humankind,
  241. and even life itself.
  242. The last time something
    so important has happened
  243. was maybe 3.5 billion years ago,
    when life was invented.
  244. A new type of life is going to emerge
    from our little planet
  245. and it's going to colonize
    and transform the entire universe.
  246. The universe is still young:
    it's only 13.8 billion years old,
  247. it's going to become much older than that,
    many times older than that.
  248. So there's plenty of time
    to reach all of it,
  249. or all of the visible parts,
  250. totally within the limits
    of light speed and physics.
  251. A new type of life is going
    to make the universe intelligent.
  252. Now, of course, we are not going to remain
    the crown of creation, of course not.
  253. But there is still beauty
  254. in seeing yourself
    as part of a grander process
  255. that leads the cosmos
  256. from low complexity
    towards higher complexity.
  257. It's a privilege to live at a time
  258. where we can witness
    the beginnings of that
  259. and where we can contribute
    something to that.
  260. Thank you for your patience.
  261. (Applause)