Google's software first tries to pick out the individual parts of speech – the different types of vowels and consonants that make up words. That's one layer of the neural network. Then it uses that information to build more sophisticated guesses, each layer of these connections drives it closer to figuring out what's being said.
Neural network algorithms can be used to analyze images too. "What you want to do is find little pieces of structure in the pixels, like for example like an edge in the image," says Hinton. "You might have a layer of feature-detectors that detect things like little edges. And then once you've done that you have another layer of feature detectors that detect little combinations of edges like maybe corners. And once you've done that, you have another layer and so on."
Neural networks promised to do something like this back in the 1980s, but getting things to actually work at the multiple levels of analysis that Hinton describes was difficult.
But in 2006, there were two big changes. First, Hinton and his team figured out a better way to map out deep neural networks – networks that make many different layers of connections. Second, low-cost graphical processing units came along, giving the academics had a much cheaper and faster way to do the billions of calculations they needed. "It made a huge difference because it suddenly made things go 30 times as fast," says Hinton.
Today, neural network algorithms are starting to creep into voice recognition and imaging software, but Hinton sees them being used anywhere someone needs to make a prediction. In November, a University of Toronto team used neural networks to predict how drug molecules might behave in the real world.
Jeff Dean says that Google is now using neural network algorithms in a variety of products – some experimental, some not – but nothing is as far along as the Jelly Bean speech recognition software. "There are obvious tie-ins for image search," he says. "You'd like to be able to use the pixels of the image and then identify what object that is." Google Street View could use neural network algorithms to tell the difference between different kinds of objects it photographs – a house and a license plate, for example.
And lest you think this may not matter to regular people, take note. Last year Google researchers, including Dean, built a neural network program that taught itself to identify cats on YouTube.
Microsoft and IBM are studying neural networks too. In October, Microsoft Chief Research Officer Rick Rashid showed a live demonstration of Microsoft's neural network-based voice processing software in Tianjin, China. In the demo, Rashid spoke in English and paused after each phrase. To the audience's delight, Microsoft's software simultaneously translated what he was saying and then spoke it back to the audience in Chinese. The software even adjusted its intonation to make itself sound like Rashid's voice.
"There's much work to be done in this area," he said. "But this technology is very promising, and we hope in a few years that we'll be able to break down the language barriers between people. Personally, I think this is going to lead to a better world."
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4