Most Active Stories
- Broward School Board Suspends Teacher Who Used Slur Against Muslim Student
- An Idea To Mitigate Rising Seas In Miami Beach: Lift The Entire City
- Which One Is Better: Miami Or Miami Beach?
- How An Ethnic Slur Spurred A Broward Father's Activism
- Stalin Stupor: Why Venezuela Keeps Getting Ranked "Most Miserable" In 2015
All Tech Considered
Thu February 20, 2014
Deep Learning: Teaching Computers To Tell Things Apart
Originally published on Thu February 20, 2014 9:03 pm
WhatsApp may be Facebook's latest prize, but it's not the company's most ambitious investment. In recent months, the social networking giant has begun funding something potentially far more revolutionary: artificial intelligence.
And it's not alone. Google and as of last week, Netflix, are all getting into a new kind of AI known as deep learning. Deep learning programs are able to perceive their world in a way unlike any other computer program, and these companies hope that deep learning programs will one day be able to sort your photos, recognize your voice, and do a whole lot more.
Computers can already recognize some things. Take the scanner at the supermarket. You swipe your block of cheese, and the scanner reads the bar code. Not that you would ever do this, but try scanning something it doesn't expect — like a cat — and the machine will be hopelessly confused.
Yann LeCun, a computer scientist at New York University, is developing deep learning programs. He's trying to create a program that can recognize things without bar codes. Things it hasn't seen before. Show his program a cat. And the program sees: nose, whiskers, ears, and takes a guess.
The first time, it might guess dog. But the human user can correct it. And from that correction, it figures out what it did wrong. "Then it adjusts, so that next time around it will give you a better answer," LeCun says.
This is the key thing about LeCun's program: It learns to recognize things on its own, without bar codes or labels. And it gets really good. One of his students has built a program that can tell cat from dog 99 percent of the time.
Deep learning "really doesn't look like a computer program," says Gary Marcus a psychologist and AI expert at New York University who doesn't work with LeCun.
He says ordinary computer code is written in very strict logical steps. "But what you'll see in deep learning is something different; you don't have a lot of instructions that say: 'If one thing is true do this other thing,' " he says.
Instead of linear logic, deep learning is based on theories of how the human brain works. The program is made of tangled layers of interconnected nodes. It learns by rearranging connections between nodes after each new experience.
New supercomputers and giant databases of images make deep learning programs hugely powerful. LeCun exposed one of his programs to a database filled with 1.5 million random pictures. After going through every picture in the database, the program could identify pretty much everything. It could identify specific breeds of dogs and monkeys. And it could do it even if the monkey was in a tree, or the dog was under a bed.
LeCun's program was so good it got Facebook's attention. Late last year, he got a call from Facebook CEO Mark Zuckerberg. LeCun says the conversation was deep. "A lot of it was about philosophy really: How do we go about building machines? Is it a good idea to try and build intelligent machines to help Facebook's business?" he says.
At the end of the call, LeCun got a new job: He's now director of AI research at Facebook.
LeCun says deep learning will soon change the way Facebook and other companies handle your images. It may be possible for Facebook's search engine to know the face of a friend, or to organize images that are taken in the same place or with the same people. "You're going to have similar things happening for video," he adds.
NYU's Marcus says there are limits to what deep learning can do. It can hear words, for example, but it can't really understand what a sentence means. "Google, I think, would like to build search engines that understand the English language, and deep learning's not really powerful enough to do that," he says. "But it's powerful enough to help with a lot of things."
AUDIE CORNISH, HOST:
From NPR News, this is ALL THINGS CONSIDERED. I'm Audie Cornish.
ROBERT SIEGEL, HOST:
And I'm Robert Siegel.
Facebook has just offered billions for the messaging company WhatsApp. It's the latest in a string of record acquisitions by Silicon Valley tech giants. In recent months, Facebook has also begun quietly investing in research.
NPR's Geoff Brumfiel reports on its new venture into artificial intelligence.
GEOFF BRUMFIEL, BYLINE: Until recently, Yann LeCun was just another computer scientist at New York University. He specialized in trying to get computers to think more like people. Most of us know it as artificial intelligence.
YANN LECUN: It's an obsession that goes back a very long time.
BRUMFIEL: The goal of AI is to make computers that interact with people as if they were people, too. But at the moment, LeCun is focused on a smaller part of that goal: getting computers to recognize things the way we do. Computers already recognize some things. Take the scanner at the supermarket. You swipe your block of cheese, and the scanner reads the bar code
(SOUNDBITE OF BEEP)
COMPUTERIZED VOICE: 6.99.
BRUMFIEL: But when you think about it, that's a pretty limited kind of recognition. Try scanning something that doesn't have a bar code. I don't know - a cat.
(SOUNDBITE OF MEOW)
COMPUTERIZED VOICE: Unexpected item in bagging area.
BRUMFIEL: Yann LeCun is trying to create a computer program that can recognize things without bar codes, things it hasn't seen before. Show his program a cat...
(SOUNDBITE OF MEOW)
BRUMFIEL: ...and the program sees nose, whiskers, ears; and takes a guess. The first time it might guess dog. But you tell it wrong; cat. It figures out what it did wrong.
LECUN: And then it adjusts so that next time around that it sees the same image, it will give you a better answer.
BRUMFIEL: This is the key thing about LeCun's program. It learns to recognize things on its own, without bar codes or labels. It's called deep learning. And it's radically different from what's come before.
GARY MARCUS: It really doesn't look like a computer program.
BRUMFIEL: Gary Marcus is a psychologist at New York University who doesn't work with LeCun. He says ordinary computer code is written in very strict, logical steps. If the scan shows Cheerios, then charge 4.49.
MARCUS: But what you'll see in deep learning is something different. You don't have a lot of instructions that say: If one thing is true, do this other thing.
BRUMFIEL: Instead of linear logic, deep learning is loosely based on theories of how the human brain works. The program is a sprawling tangle of connections. It learns by rearranging those connections after each new experience. Now, Marcus says the idea of a self-reorganizing computer isn't entirely new. It's actually based on a machine from the 1950s, called the Perceptron.
But by using new supercomputers and giant databases of images, deep learning programs are getting way past supermarket scanner territory. Yann LeCun exposed one of his programs to a huge database filled with random pictures.
LECUN: On the order of 1.5 million images.
BRUMFIEL: After going through every picture in the database, the program could identify pretty much everything, things I've never even heard of.
LECUN: Do you know what a langur is?
BRUMFIEL: Not a clue.
LECUN: OK, it's a particular type of monkey. And you show it a picture of a langur, and it'll tell you it's a langur. Or a Samoyed - do you know what a Samoyed is? - or Samoyed; I don't know how you pronounce it. That's actually a breed of dog,and it will tell you it's a Samoyed. It won't tell you, it's a dog. It will tell you, that's this breed of dog.
BRUMFIEL: What's more, the computer didn't have to have a close-up of the dog. It could identify one even if it was hiding under a chair. LeCun's program was so good, it got Facebook's attention. Late last year, he got a call from Facebook CEO Mark Zuckerberg. LeCun says the conversation was deep.
LECUN: A lot of it was about philosophy, really. How do we go about building intelligent machines? Is it a good idea to try to build intelligent machines to help Facebook's business?
BRUMFIEL: At the end of the call, LeCun got a new job. He's now director of AI research at Facebook. Google has also hired deep learning researchers. But Gary Marcus says there are limits to what deep learning can do. It can hear words, for example, but it can't really understand what a sentence means.
MARCUS: Google, I think, would like to build search engines that really understand the English language, and deep learning is not really powerful enough to do that. But it's powerful enough to help with a lot of things.
BRUMFIEL: Google and Facebook have big plans for deep learning. In the next few years, they will use these programs to better recognize your voice, the people in your pictures, and a whole lot more.
Geoff Brumfiel, NPR News. Transcript provided by NPR, Copyright NPR.