WEBVTT

00:00:04.108 --> 00:00:06.769
♪ (Music fades in) ♪

00:00:15.092 --> 00:00:19.394
(Chirping)

00:00:24.559 --> 00:00:30.051
(Vocalizations, different languages)

00:00:32.488 --> 00:00:35.487
(Talking overlaps in background)

00:00:36.347 --> 00:00:37.786
(Laughing)

00:00:38.486 --> 00:00:43.801
(Computerized beeping)

00:00:45.241 --> 00:00:46.556
(Beep)

00:00:49.116 --> 00:00:54.431
(Man) We come into this world with the
innate ability to learn to interact

00:00:54.431 --> 00:00:57.320
with other sentient beings.

00:00:58.060 --> 00:00:59.767
(Child vocalizing)

00:01:00.116 --> 00:01:02.542
(Man) Suppose you had to interact with
other people by writing little messages.

00:01:03.452 --> 00:01:04.768
(Child vocalizes)

00:01:05.258 --> 00:01:07.396
(Man) It'd be a real pain.

00:01:07.396 --> 00:01:09.486
(Man) And that's how we interact
with computers.

00:01:09.490 --> 00:01:11.647
It's much easier just to talk to them...
just so much easier...

00:01:12.137 --> 00:01:13.314
(Child vocalizes)

00:01:13.804 --> 00:01:16.491
(Man) If the computers could understand
what we're saying.

00:01:17.814 --> 00:01:20.342
For that, you need really 
good speech recognition.

00:01:20.547 --> 00:01:23.859
(Narrator) The first speech recognition
system was developed by Bell Laboratories

00:01:23.859 --> 00:01:28.108
in 1952. It could only recognize
numbers spoken by one person.

00:01:28.108 --> 00:01:31.500
In the 1970s, Carnegie-Mellon
came out with the Harpy System.

00:01:31.500 --> 00:01:36.522
This was able to recognize over
1,000 words and different pronunciations

00:01:36.522 --> 00:01:39.563
(Narrator) of the same word.
- (Man) Tomato - (Woman) Tomato

00:01:39.563 --> 00:01:42.543
(Narrator) Speech recognition continued
in the 80s with the introduction of the

00:01:42.543 --> 00:01:45.576
Hidden Markov Model, which
used a more mathematical approach

00:01:45.576 --> 00:01:50.193
to analyzing sound waves that led to
many breakthroughs we have today.

00:01:50.197 --> 00:01:52.904
You're taking in very raw audio wave forms

00:01:52.904 --> 00:01:54.623
like you get through a microphone

00:01:54.623 --> 00:01:55.858
on your phone

00:01:55.858 --> 00:01:56.831
or whatever...

00:01:56.966 --> 00:02:02.446
(Woman) We chop it into small pieces
and it tries to identify which phoneme

00:02:02.446 --> 00:02:05.498
was spoken in that piece of speech.

00:02:05.498 --> 00:02:09.450
- Phoneme is a primitive unit for
expressing words.

00:02:09.530 --> 00:02:13.924
(voicing phonemes shown above)

00:02:14.635 --> 00:02:19.757
And then it stitches those together
into likely words like Palo Alto.

00:02:19.757 --> 00:02:23.593
- Speech recognition today is good at
transcribing what you've said...

00:02:23.593 --> 00:02:25.355
(Man, to phone) What's the weather
like in Topeka?

00:02:25.355 --> 00:02:30.122
(Man) You can talk about travels, your
contacts, like, "Where can I get pizza?"

00:02:30.122 --> 00:02:32.072
(Phone) Here are the listings for Pizza.

00:02:32.072 --> 00:02:34.281
(Man) "How tall is the Eiffel Tower?"
(Phone) The Eiffel Tower is ...

00:02:34.326 --> 00:02:36.912
(Woman) We've made tremendous
improvements very quickly.

00:02:36.912 --> 00:02:39.212
(Man, to phone) Who is the 21st
President of the United States?

00:02:39.560 --> 00:02:42.360
(Phone beeps)
(Phone) Chester A. Arthur was the 21st...

00:02:42.360 --> 00:02:44.397
(Man, to phone) Okay, Google,
where is he from?

00:02:44.397 --> 00:02:47.303
(Man) Years ago, you had to be an engineer
to interact with computers.

00:02:47.953 --> 00:02:50.023
Today, everybody can interact.

00:02:50.234 --> 00:02:53.772
- One thing still in its
infancy is understanding.

00:02:53.772 --> 00:02:56.412
- We need a far more sophisticated
language understanding model

00:02:56.412 --> 00:02:58.592
that understands what the sentence means.

00:02:58.592 --> 00:03:00.902
We're still a very long way from that.

00:03:01.187 --> 00:03:02.511
(Beeping)

00:03:03.501 --> 00:03:06.912
♪ (Soft background music) ♪

00:03:07.745 --> 00:03:11.625
(Woman) Our ability to use language is one
of the things that helps us have culture.

00:03:13.467 --> 00:03:18.649
It's one of the things that helps
us pass on traditions across generations.

00:03:19.516 --> 00:03:25.975
Figuring out how the system of language
works, even though it seems easy,

00:03:25.982 --> 00:03:32.666
turns out to be very hard, but is one that
every baby understands by 2 years old.

00:03:32.666 --> 00:03:37.529
(Girl) There's two of them.
(Woman) There's two Ls, yeah (spells word)

00:03:38.466 --> 00:03:41.045
- Language is extremely complex
and sophisticated...

00:03:41.086 --> 00:03:42.445
- From the semantics

00:03:42.445 --> 00:03:43.819
- (Man in chair) Ironies...
- (Woman) Strong accents...

00:03:43.819 --> 00:03:44.887
- (Man) Facial expressions...

00:03:44.887 --> 00:03:47.445
- Human emotions, because that's
part of how we communicate.

00:03:47.445 --> 00:03:48.846
- Humor...

00:03:48.846 --> 00:03:51.573
(Aside) Do I have to be careful
not to offend the dinosaur?

00:03:51.573 --> 00:03:54.187
- Language has so many different
layers and that's why it's

00:03:54.187 --> 00:03:56.126
such a difficult problem.

00:03:56.126 --> 00:03:59.214
(Man) The present human brain
and the learning algorithms in it

00:03:59.214 --> 00:04:01.997
are far, far better at things like
language understanding

00:04:02.179 --> 00:04:04.781
and they're still a lot better
at pun recognition.

00:04:05.032 --> 00:04:09.379
- Whether or not we replicate exactly
what the brain does, to understand

00:04:09.379 --> 00:04:14.194
language and speech, is still a question.

00:04:15.125 --> 00:04:16.951
(Beeping)

00:04:17.506 --> 00:04:23.306
(Man) For many years, we believed that
neural networks should work better than

00:04:23.306 --> 00:04:26.914
the dumb existing technology that's
basically just "table look-up."

00:04:27.443 --> 00:04:33.481
Then, in 2009, two of my students
(with some help from me) got it

00:04:33.481 --> 00:04:36.849
working better. The first time it was
just a little better.

00:04:36.849 --> 00:04:40.165
But it was obvious that this could be
improved to work much better.

00:04:40.485 --> 00:04:44.442
(Man) The brain has this system of neurons
all computing in parallel.

00:04:44.537 --> 00:04:48.882
All knowledge in the brain is in the
strength of connection between neurons.

00:04:49.212 --> 00:04:53.219
What I mean by "neural net" is something
that is simulated on a conventional

00:04:53.219 --> 00:04:58.624
computer, but is designed to work in
roughly the same ways as the brain.

00:04:59.804 --> 00:05:03.954
Until quite recently, people got features
by hand engineering them.

00:05:04.203 --> 00:05:08.600
They looked at sine waves and did fourier
analysis and tried to figure out

00:05:08.600 --> 00:05:11.925
what features they should feed to the
pattern recognition system.

00:05:12.045 --> 00:05:14.517
The thing about neural networks is that
they learn their own features.

00:05:14.569 --> 00:05:19.552
In particular, they can learn features
and features of features, etc,

00:05:21.065 --> 00:05:23.770
and that's lead to huge improvement
in speech recognition.

00:05:23.860 --> 00:05:26.710
- But you can also use them for language
understanding tasks.

00:05:26.842 --> 00:05:32.690
How you do this is to represent words
in very high-dimensional spaces.

00:05:32.724 --> 00:05:36.200
- (Man) We can now deal with analogies
where a word is represented as a list

00:05:36.200 --> 00:05:37.567
of numbers.

00:05:37.748 --> 00:05:44.292
For example, if I take 100 numbers that
represent "Paris," and I subtract from it

00:05:44.292 --> 00:05:49.669
"France" and add "Italy," if I look
at the numbers I have, the closest

00:05:49.669 --> 00:05:53.000
thing is a list of numbers that
represents "Rome."

00:05:53.155 --> 00:05:58.038
By first converting words into numbers,
using a neural net, you can actually

00:05:58.038 --> 00:06:00.772
do this analogical reasoning.

00:06:01.194 --> 00:06:06.286
I predict that, in the next five years, it
will be clear that these neural networks

00:06:06.286 --> 00:06:10.832
with new learning algorithms will give us
much better language understanding.

00:06:13.476 --> 00:06:19.079
(Woman) When we started out, we thought
things like chess or mathematics or logic

00:06:19.079 --> 00:06:21.303
would be things that were really hard.

00:06:21.303 --> 00:06:26.177
They're not that hard. We ended up with
a machine that played as well as

00:06:26.177 --> 00:06:28.489
a Grand Master at chess.

00:06:28.489 --> 00:06:33.092
What we thought would be easy for
a computer system, like language,

00:06:33.092 --> 00:06:36.730
has turned out incredibly hard.

00:06:36.730 --> 00:06:41.692
(Man) I can't even imagine the moment of
success quite yet because there are so many

00:06:41.692 --> 00:06:46.534
pieces of this puzzle that are unsolved,
both from a science point of view

00:06:46.534 --> 00:06:50.573
as well as a technical implementation
point of view.

00:06:50.573 --> 00:06:52.372
There are a lot of unknowns.

00:06:52.372 --> 00:06:56.257
(Woman) Those are the great revolutions.
Not just what we fiddle with what

00:06:56.257 --> 00:07:00.230
we already know, but when we discover
something completely new and unexpected.

00:07:00.230 --> 00:07:03.252
(Man) Once you are in the area of

00:07:03.252 --> 00:07:08.534
human-level performance, 
that will be pretty remarkable.

00:07:12.760 --> 00:07:14.361
(Beep)