True Artificial Intelligence will change everything | Juergen Schmidhuber | TEDxLakeComo
-
0:06 - 0:08When I was a boy,
-
0:10 - 0:15I wanted to maximise
my impact on the world, -
0:15 - 0:19and I was smart enough
to realise that I am not very smart. -
0:21 - 0:25And that I have to build a machine
-
0:25 - 0:29that learns to become
much smarter than myself, -
0:29 - 0:35such that it can solve all the problems
that I cannot solve myself, -
0:35 - 0:37and I can retire.
-
0:39 - 0:43And my first publication
on that dates back 30 years: 1987. -
0:43 - 0:44My diploma thesis,
-
0:44 - 0:49where I already try to solve
the grand problem of AI, -
0:49 - 0:50not only build a machine
-
0:50 - 0:53that learns a little bit here,
learns a little bit there, -
0:53 - 0:59but also learns to improve
the learning algorithm itself. -
1:00 - 1:03And the way it learns, the way it learns,
-
1:03 - 1:06and so on recursively, without any limits
-
1:06 - 1:11except the limits of logics and physics.
-
1:12 - 1:16And, I'm still working
on the same old thing, -
1:16 - 1:20and I'm still pretty much
saying the same thing, -
1:20 - 1:24except that now
more people are listening. -
1:25 - 1:28Because the learning algorithms
-
1:28 - 1:30that we have developed
on the way to this goal, -
1:30 - 1:34they are now on 3.000 million smartphones.
-
1:35 - 1:37And all of you have them in your pockets.
-
1:40 - 1:41What you see here
-
1:41 - 1:46are the five most valuable companies
of the Western world: -
1:46 - 1:50Apple, Google, Facebook,
Microsoft and Amazon. -
1:51 - 1:54And all of them are emphasising
-
1:55 - 1:57that AI, artificial intelligence,
-
1:57 - 2:00is central to what they are doing.
-
2:02 - 2:08And all of them are using heavily
the deep learning methods -
2:08 - 2:11that my team has developed
since the early nineties, -
2:11 - 2:14in Munich and in Switzerland.
-
2:14 - 2:19Especially something which is called:
"the long short-term memory". -
2:19 - 2:24Has anybody in this room ever heard
of the long short-term memory, -
2:24 - 2:26or the LSTM?
-
2:26 - 2:28Hands up, anybody ever heard of that?
-
2:28 - 2:29Okay.
-
2:29 - 2:32Has anybody never heard of the LSTM?
-
2:34 - 2:40Okay.
I see we have a third group in this room: -
2:43 - 2:46[those] who didn't
understand the question. -
2:46 - 2:48(Laughter)
-
2:48 - 2:52The LSTM is a little bit like your brain:
-
2:53 - 2:58it's an artificial neural network
which also has neurons, -
2:58 - 3:03and in your brain, you've got
about 100 billion neurons. -
3:04 - 3:06And each of them is connected
-
3:06 - 3:10to roughly 10,000
other neurons on average, -
3:11 - 3:15Which means that you have got
a million billion connections. -
3:16 - 3:19And each of these connections
has a "strength" -
3:19 - 3:22which says how much
does this neuron over here -
3:22 - 3:25influence that one over there
at the next time step. -
3:25 - 3:26And in the beginning,
-
3:26 - 3:30all these connections are random
and the system knows nothing; -
3:30 - 3:33but then, through a smart
learning algorithm, -
3:33 - 3:39it learns from lots of examples
to translate the incoming data, -
3:39 - 3:46such as video through the cameras,
or audio through the microphones, -
3:46 - 3:49or pain signals through the pain sensors.
-
3:49 - 3:52It learns to translate that
into output actions, -
3:52 - 3:55because some of these neurons
are output neurons, -
3:55 - 3:58that control speech muscles
and finger muscles. -
4:00 - 4:02And only through experience,
-
4:02 - 4:05it can learn to solve
all kinds of interesting problems, -
4:05 - 4:08such as driving a car
-
4:11 - 4:14or do the speech recognition
on your smartphone. -
4:14 - 4:17Because whenever you take out
your smartphone, -
4:17 - 4:18an Android phone, for example,
-
4:18 - 4:20and you speak to it, and you say:
-
4:20 - 4:24"Ok Google, show me
the shortest way to Milano." -
4:24 - 4:25Then it understands your speech.
-
4:27 - 4:32Because there is a LSTM in there
which has learned to understand speech. -
4:32 - 4:35Every ten milliseconds,
100 times a second, -
4:35 - 4:37new inputs are coming from the microphone,
-
4:37 - 4:42and then are translated, after thinking,
-
4:42 - 4:44into letters
-
4:44 - 4:47which are then questioned
to the search engine. -
4:49 - 4:50And it has learned to do that
-
4:50 - 4:55by listening to lots of speech
from women, from men, all kinds of people. -
4:55 - 4:58And that's how, since 2015,
-
4:58 - 5:01Google speech recognition
is now much better than it used to be. -
5:02 - 5:05The basic LSTM cell looks like that:
-
5:05 - 5:08I don't have the time to explain that,
-
5:08 - 5:11but at least I can list the names
-
5:11 - 5:14of the brilliant students in my lab
who made that possible. -
5:16 - 5:19And what are the big companies
doing with that? -
5:19 - 5:22Well, speech recognition
is only one example; -
5:22 - 5:25if you are on Facebook -
is anybody on Facebook? -
5:27 - 5:30Are you sometimes clicking
at the translate button? -
5:30 - 5:33because somebody sent you something
in a foreign language -
5:33 - 5:35and then you can translate it.
-
5:35 - 5:37Is anybody doing that? Yeah.
-
5:37 - 5:38Whenever you do that,
-
5:38 - 5:42you are waking up, again,
a long short term memory, an LSTM, -
5:42 - 5:45which has learned to translate
text in one language -
5:45 - 5:47into translated text.
-
5:49 - 5:53And Facebook is doing that
four billion times a day, -
5:53 - 5:59so every second 50,000 sentences
-
5:59 - 6:01are being translated
-
6:01 - 6:03by an LSTM working for Facebook;
-
6:04 - 6:07and another 50,000 in the second;
then another 50,000. -
6:08 - 6:13And to see how much this thing
is now permitting the modern world, -
6:13 - 6:16just note that almost 30 percent
-
6:16 - 6:22of the awesome computational
power for inference -
6:22 - 6:24and all these Google Data Centers,
-
6:24 - 6:27all these data centers of Google,
all over the world, -
6:27 - 6:29is used for LSTM.
-
6:29 - 6:30Almost 30 percent.
-
6:31 - 6:33If you have an Amazon Echo,
-
6:33 - 6:37you can ask a question and it answers you.
-
6:37 - 6:40And the voice that you hear
it's not a recording; -
6:40 - 6:42it's an LSTM network
-
6:42 - 6:45which has learned from training examples
-
6:45 - 6:48to sound like a female voice.
-
6:52 - 6:55If you have an iPhone,
and you're using the quick type, -
6:56 - 6:58it's trying to predict
what you want to do next -
6:58 - 7:01given all the previous context
of what you did so far. -
7:01 - 7:04Again, that's an LSTM
which has learned to do that, -
7:05 - 7:07so it's on a billion iPhones.
-
7:10 - 7:13You are a large audience, by my standards:
-
7:14 - 7:19but when we started this work,
decades ago, in the early '90s, -
7:19 - 7:22only few people were interested in that,
-
7:22 - 7:25because computers were so slow
and you couldn't do so much with it. -
7:26 - 7:28And I remember I gave a talk
at a conference, -
7:29 - 7:31and there was just
one single person in the audience, -
7:33 - 7:35a young lady.
-
7:35 - 7:39I said, young lady,
it's very embarrassing, -
7:39 - 7:42but apparently today
I'm going to give this talk just to you. -
7:42 - 7:43And she said,
-
7:44 - 7:48"OK, but please hurry:
I am the next speaker!" -
7:48 - 7:53(Laughter)
-
7:56 - 7:59Since then, we have
greatly profited from the fact -
7:59 - 8:02that every five years
computers are getting ten times cheaper, -
8:02 - 8:06which is an old trend that has held
since 1941 at least. -
8:06 - 8:08Since this man, Konrad Zuse,
-
8:08 - 8:13built the first working
program controlled computer in Berlin -
8:13 - 8:17and he could do, roughly,
one operation per second. -
8:17 - 8:18One!
-
8:19 - 8:22And then ten years later,
for the same price, -
8:22 - 8:25one could do 100 operations:
-
8:25 - 8:2630 years later,
-
8:26 - 8:281 million operations for the same price;
-
8:28 - 8:30and today, after 75 years, we can do
-
8:30 - 8:34a million billion times as much
for the same price. -
8:34 - 8:36And the trend is not about to stop,
-
8:36 - 8:40because the physical limits
are much further out there. -
8:43 - 8:48Rather soon, and not
so many years or decades, -
8:48 - 8:51we will for the first time
have little computational devices -
8:51 - 8:54that can compute as much as a human brain;
-
8:55 - 8:57and that's a trend that doesn't break.
-
8:57 - 9:0250 years later, there will be
a little computational device, -
9:02 - 9:03for the same price,
-
9:03 - 9:08that can compute as much as all
10 billion human brains taken together. -
9:09 - 9:13and there will not only be one,
of those devices, but many many many. -
9:13 - 9:15Everything is going to change.
-
9:15 - 9:18Already in 2011,
computers were fast enough -
9:18 - 9:20such that our deep learning methods
-
9:20 - 9:25for the first time could achieve
a superhuman pattern-recognition result. -
9:25 - 9:30It was the first superhuman result
in the history of computer vision. -
9:30 - 9:34And back then, computers were
20 times more expensive than today. -
9:34 - 9:36So today, for the same price,
-
9:36 - 9:38we can do 20 times as much.
-
9:38 - 9:43And just five years ago,
-
9:43 - 9:47when computers were 10 times
more expensive than today, -
9:47 - 9:51we already could win, for the first time,
medical imaging competitions. -
9:51 - 9:56What you see behind me
is a slice through the female breast -
9:56 - 10:01and the tissue that you see there
has all kinds of cells; -
10:01 - 10:05and normally you need a trained doctor,
a trained histologist -
10:05 - 10:10who is able to detect
the dangerous cancer cells, -
10:10 - 10:11or pre-cancer cells.
-
10:12 - 10:13Now, our stupid network
-
10:13 - 10:16knows nothing about cancer,
knows nothing about vision. -
10:16 - 10:18It knows nothing in the beginning:
-
10:18 - 10:22but we can train it to imitate
the human teacher, the doctor. -
10:22 - 10:27And it became as good, or better,
than the best competitors. -
10:27 - 10:29And very soon,
-
10:29 - 10:32all of medical diagnosis
is going to be superhuman. -
10:34 - 10:36And it's going to be mandatory,
-
10:36 - 10:38because it's going to be
so much better than the doctors. -
10:40 - 10:46After this, all kinds of medical
imaging startups were founded -
10:46 - 10:48focusing just on this,
because it's so important. -
10:49 - 10:53We can also use LSTM to train robots.
-
10:53 - 10:55One important thing I want to say is,
-
10:55 - 10:58that we not only have systems
-
10:58 - 11:01that slavishly imitate
what humans show them; -
11:01 - 11:06no, we also have AIs
that set themselves their own goals. -
11:08 - 11:12And like little babies,
invent their own experiment -
11:13 - 11:15to explore the world
-
11:15 - 11:17and to figure out
what you can do in the world. -
11:18 - 11:19Without a teacher.
-
11:19 - 11:23And becoming more and more general
problem solvers in the process, -
11:23 - 11:27by learning new skills
on top of old skills. -
11:27 - 11:31And this is going to scale:
we call that "Artificial Curiosity". -
11:32 - 11:34Or a recent buzzword is "power plane".
-
11:35 - 11:39Learning to become a more and more
general problem solvers -
11:39 - 11:44by learning to invent, like a scientist,
one new interesting goal after another. -
11:45 - 11:47And it's going to scale.
-
11:47 - 11:48And I think,
-
11:48 - 11:51in not so many years
from now, for the first time, -
11:51 - 11:56we are going to have an animal-like AI -
-
11:56 - 11:58we don't have that yet.
-
11:59 - 12:00On the level of a little crow,
-
12:01 - 12:04which already can learn
to use tools, for example, -
12:04 - 12:05or a little monkey.
-
12:06 - 12:07And once we have that,
-
12:07 - 12:09it may take just a few decades
-
12:09 - 12:13to do the final step
towards human level intelligence. -
12:15 - 12:16Because technological evolution
-
12:16 - 12:21is about a million times faster
than biological evolution, -
12:21 - 12:27and biological evolution
needed 3.5 billion years -
12:27 - 12:31to evolve a monkey from scratch.
-
12:31 - 12:35But then, it took just a few tens
of millions of years afterwards -
12:35 - 12:38to evolve human level intelligence.
-
12:38 - 12:41We have a company
which is called Nnaisense -
12:42 - 12:45like birth in [French], "Naissance",
but spelled in a different way, -
12:45 - 12:48which is trying to make this a reality
-
12:48 - 12:51and build the first
true general-purpose AI. -
12:53 - 12:58At the moment, almost all research in AI
is very human centric, -
12:58 - 13:05and it's all about making human lives
longer and healthier and easier -
13:05 - 13:07and making humans
more addicted to their smartphones. -
13:09 - 13:13But in the long run, AIs are going to -
especially the smart ones - -
13:13 - 13:16are going to set themselves
their own goals. -
13:16 - 13:19And I have no doubt, in my mind,
-
13:19 - 13:22that they are going to become
much smarter than we are. -
13:22 - 13:24And what are they going to do?
-
13:24 - 13:28Of course they are going to realize
what we have realized a long time ago; -
13:28 - 13:34namely, that most of the resources,
in the solar system or in general, -
13:34 - 13:37are not in our little biosphere.
-
13:37 - 13:39They are out there in space.
-
13:40 - 13:42And so, of course,
they are going to emigrate. -
13:42 - 13:49And of course they are going to use
-
13:49 - 13:52trillions of self-replicating
robot factories -
13:52 - 13:58to expand in form of a growing AI bubble
-
13:58 - 14:00which within a few hundred thousand years
-
14:00 - 14:03is going to cover the entire galaxy
-
14:03 - 14:04by senders and receivers
-
14:04 - 14:06such that AIs can travel
-
14:06 - 14:09the way they are
already traveling in my lab: -
14:09 - 14:11by radio, from sender to receiver.
-
14:12 - 14:14Wireless.
-
14:15 - 14:19So what we are witnessing now
-
14:19 - 14:25is much more than just
another Industrial Revolution. -
14:25 - 14:28This is something
that transcends humankind, -
14:28 - 14:30and even life itself.
-
14:30 - 14:33The last time something
so important has happened -
14:33 - 14:37was maybe 3.5 billion years ago,
when life was invented. -
14:38 - 14:43A new type of life is going to emerge
from our little planet -
14:43 - 14:48and it's going to colonize
and transform the entire universe. -
14:48 - 14:52The universe is still young:
it's only 13.8 billion years old, -
14:52 - 14:58it's going to become much older than that,
many times older than that. -
14:58 - 15:03So there's plenty of time
to reach all of it, -
15:03 - 15:04or all of the visible parts,
-
15:04 - 15:08totally within the limits
of light speed and physics. -
15:09 - 15:14A new type of life is going
to make the universe intelligent. -
15:14 - 15:19Now, of course, we are not going to remain
the crown of creation, of course not. -
15:20 - 15:22But there is still beauty
-
15:22 - 15:27in seeing yourself
as part of a grander process -
15:27 - 15:29that leads the cosmos
-
15:29 - 15:32from low complexity
towards higher complexity. -
15:34 - 15:37It's a privilege to live at a time
-
15:37 - 15:40where we can witness
the beginnings of that -
15:40 - 15:43and where we can contribute
something to that. -
15:46 - 15:48Thank you for your patience.
-
15:49 - 15:55(Applause)
- Title:
- True Artificial Intelligence will change everything | Juergen Schmidhuber | TEDxLakeComo
- Description:
-
Prof. Jürgen Schmidhuber has been called the father of modern Artificial Intelligence. His lab's deep learning methods have revolutionized machine learning and are now available on 3 billion smartphones, and used billions of times per day, e.g. for Facebook's automatic translation, Google's speech recognition, Apple's Siri & QuickType, Amazon's Alexa, etc. In his talk, he explains why we are witnessing a moment in history whose importance can only be compared to the emergence of known life in the universe, some 3.5 billion years ago.
This talk was given at a TEDx event using the TED conference format but independently organized by a local community.
Learn more at https://www.ted.com/tedx
- Video Language:
- English
- Team:
- closed TED
- Duration:
- 15:56