GEOFFREY HINTON: We come into this world with the innate abilities to learn to interact with other sentient beings. Suppose you had to interact with other people by writing little messages to them. It’d be a real pain. And that’s how we interact with computers. It’s much easier just to talk to them. It’s just so much easier if the computers could understand what we’re saying. And for that, you need really good speech recognition. NARRATOR: The first speech recognition system was developed by Bell Laboratories in 1952. It could only recognize numbers spoken by one person. In the 1970s, Carnegie Mellon came out with the Harpy system, which was able to recognize over 1,000 words and could recognize different pronunciations of the same word.
MALE COMPUTER VOICE: Tomato. FEMALE COMPUTER VOICE: Tomato. NARRATOR: Speech recognition continued in the ’80s with the introduction of the hidden Markov model, which used a more mathematical approach to analyzing sound waves and led to many of the breakthroughs we have today. JEFF DEAN: You’re taking in very raw audio waveforms. MALE SPEAKER: Like you get from a microphone on your phone or whatever. MALE COMPUTER VOICE: Cheeseburger. FRANCOISE BEAUFAYS: We chop it into small pieces and it tries to identify which phoneme was spoken in that last piece of speech. GEOFFREY HINTON: So a phoneme is a kind of primitive unit for expressing words. JEFF DEAN: JEFF DEAN: And then it will want to stitch those together into likely words like Palo Alto. RAY KURZWEIL: Speech recognition today is quite good at transcribing what you’ve said. MALE SPEAKER: What’s the weather like in Topeka? ROBERTO PIERACCINI: You can talk about travels.
You can talk about your contacts. RAY KURZWEIL: Like where can I get pizza? PHONE: Here are the listings for pizza. RAY KURZWEIL: How tall is the Eiffel Tower? PHONE: The Eiffel Tower is– FRANCOISE BEAUFAYS: We’ve made tremendous improvements very quickly. MALE SPEAKER: Who is the 21st President of the United States? PHONE: Chester A. Arthur was the 21st– MALE SPEAKER: OK, Google. Where’s he from? RAY KURZWEIL: Years ago, you had to be an engineer to interact with computers.
I mean, today, everybody can interact. ROBERTO PIERACCINI: One thing, though, that is still in the infancy is the understanding. GEOFFREY HINTON: We need a far more sophisticated language understanding model that understands what the sentence means. And we’re still a very long way from having that. ALISON GOPNIK: Our ability to use language is one of the things that helps us have culture. It’s one of the things that helps us pass on traditions from one generation to another. Figuring out about how the system of language works, even though that seems like a really easy problem, it turns out to be one that’s really hard but that every baby has cracked by the time they’re two years old.
FEMALE CHILD: There’s two L’s. FEMALE SPEAKER: There’s two L’s. Yeah. E-L-L-I and then– FEMALE CHILD: E. FEMALE SPEAKER: E. ROBERTO PIERACCINI: Language is extremely complex and sophisticated. BILL BYRNE: From the semantics– RAY KURZWEIL: Irony– FRANCOISE BEAUFAYS: Strong accents– MALE SPEAKER: Facial expressions– RAY KURZWEIL: Human emotion because that’s part of how we communicate. BILL BYRNE: Humor. RAY KURZWEIL: Do I have to be careful not to offend the dinosaur? BILL BYRNE: Language has so many different layers, and that’s why it’s such a difficult problem. GEOFFREY HINTON: At present, the human brain, and the learning algorithms in the human brain, are far, far better at things like language understanding.
And they’re still a lot better at pattern recognition. BILL BYRNE: So whether or not we replicate exactly what the brain does to understand language and to understand speech, is still a question. GEOFFREY HINTON: For many, many years, we believed that neural networks should work better than the dumb existing technology that’s basically just table lookup. And then in 2009, two of my students, with a little input from me, got it working better. And the first time it just worked a little bit better, but then it was obvious that this could be developed to something that worked much better. The brain has these gazillions of neurons all computing in parallel. And all of the knowledge in the brain is in the strength of the connection between neurons. What I mean by neural net is something that’s simulated on a conventional computer, but is designed to work in very, very roughly the same way as the brain. So until quite recently, people got features by hand engineering. They looked at sound waves, and they did Fourier analysis.
And they tried to figure out, what features should we feed to the pattern recognition system? And the thing about neural networks is they learn their own features. And in particular, they can learn features and then they can learn features of features and then they can learn features of features of features. And that’s led to a huge improvement in speech recognition. JEFF DEAN: But you can also use them for language understanding tasks.
And the way you do this is you represent words in very high dimensional spaces. GEOFFREY HINTON: We can now deal with analogies where a word is represented as a list of numbers. So for example, if I take the list of 100 numbers that represents Paris and I subtract from it France and I add to it Italy, and if I look at the numbers I’ve got, the closest thing is the list of numbers that represents Rome. So by first converting words into these numbers using a neural net, you can actually do this analogical reasoning. I predict that in the next five years, it will become clear that these big deep neural networks with the new learning algorithms are going to give us much better language understanding.
ALISON GOPNIK: When we started out, we thought that things like chess or mathematics or logic, those were going to be the things that were really hard. They’re not that hard. I mean, we can end up with a machine that actually can do chess as well as a grandmaster can play chess. The things that we thought were going to be easy for a computer system, like understanding language, those things have turned out to be incredibly hard. BILL BYRNE: I can’t even imagine the “we’ve done it” moment quite yet, just because there are so many pieces of this puzzle that are unsolved, both from a science point of view, as well as from a technical implementation point of view.
There’s a lot of unknowns. ALISON GOPNIK: Those are the great revolutions. They’re not just when we fiddle a little with what we already know, but when we discover something completely new and unexpected. JEFF DEAN: I think once you kind of are in the ballpark of human level performance, that will be pretty remarkable. .