Processor Network Unavailable Try Again Mcmillan

When Google built the latest version of its Android mobile operating system, the web giant made some big changes to the way the Bone interprets your voice commands. It installed a vox recognition system based on what'southward called a neural network – a computerized learning system that behaves much like the homo brain.

For many users, says Vincent Vanhoucke, a Google enquiry scientist who helped steer the endeavour, the results were dramatic. "Information technology kind of came every bit a surprise that we could exercise then much meliorate by just changing the model," he says.

Vanhoucke says that the voice fault charge per unit with the new version of Android – known as Jelly Edible bean – is about 25 percent lower than previous versions of the software, and that this is making people more comfy with voice commands. Today, he says, users tend to use more tongue when speaking to the phone. In other words, they act less like they're talking to a robot. "It really is changing the way that people carry."

It's simply one example of the mode neural network algorithms are changing the style our engineering science works – and they mode nosotros use it. This discipline had cooled for many years, later on spending the 1980s as one of the hottest areas of research, but now it'due south back, with Microsoft and IBM joining Google in exploring some very real applications.

When you talk to Android's phonation recognition software, the spectrogram of what yous've said is chopped upwardly and sent to eight different computers housed in Google's vast worldwide army of servers. It's then processed, using the neural network models built by Vanhoucke and his team. Google happens to be very good at breaking up big computing jobs like this and processing them very quickly, and to effigy out how to exercise this, Google turned to Jeff Dean and his squad of engineers, a grouping that's better known for reinventing the way the modern data center works.

Neural networks give researchers similar Vanhoucke a way analyzing lots and lots of patterns – in Jelly Bean's case, spectrograms of the spoken word – and and so predicting what a brand new pattern might represent. The metaphor springs from biology, where neurons in the trunk form networks with other cells that allow them to procedure signals in specialized ways. In the kind of neural network that Jelly Bean uses, Google might build up several models of how language works – one for English language voice search requests, for case – by analyzing vast swaths of real-world information.

"People accept believed for a long, long fourth dimension – partly based on what you see in the brain – that to get a good perceptual system you use multiple layers of features," says Geoffrey Hinton, a computer scientific discipline professor at the University of Toronto. "But the question is how tin you lot learn these efficiently."

Android takes a picture of the voice command and Google processes information technology using its neural network model to figure out what'southward being said.

Google's software commencement tries to pick out the individual parts of speech – the different types of vowels and consonants that make upward words. That's ane layer of the neural network. Then it uses that data to build more sophisticated guesses, each layer of these connections drives it closer to figuring out what'due south existence said.

Neural network algorithms can exist used to analyze images too. "What you want to do is observe fiddling pieces of structure in the pixels, like for example like an edge in the epitome," says Hinton. "Y'all might accept a layer of feature-detectors that find things like fiddling edges. And so in one case you've washed that you take another layer of characteristic detectors that detect little combinations of edges like maybe corners. And once you've done that, y'all take another layer and then 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.

Simply in 2006, at that place were two large 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. 2nd, low-price graphical processing units came along, giving the academics had a much cheaper and faster mode to do the billions of calculations they needed. "It fabricated a huge deviation because it suddenly made things get 30 times equally fast," says Hinton.

Today, neural network algorithms are starting to creep into vox recognition and imaging software, but Hinton sees them beingness used anywhere someone needs to brand a prediction. In Nov, a University of Toronto squad used neural networks to predict how drug molecules might carry 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 every bit the Jelly Bean speech recognition software. "There are obvious tie-ins for image search," he says. "You'd similar to be able to use the pixels of the image and so identify what object that is." Google Street View could use neural network algorithms to tell the departure between different kinds of objects information technology photographs – a house and a license plate, for case.

And lest you lot think this may not matter to regular people, take annotation. 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 Oct, Microsoft Chief Research Officer Rick Rashid showed a alive demonstration of Microsoft'south neural network-based voice processing software in Tianjin, Cathay. In the demo, Rashid spoke in English and paused afterwards each phrase. To the audience's delight, Microsoft's software simultaneously translated what he was saying so spoke it dorsum to the audience in Chinese. The software fifty-fifty adapted its intonation to brand itself audio like Rashid's voice.

"There'southward much work to exist done in this area," he said. "But this technology is very promising, and we hope in a few years that nosotros'll exist able to break downwardly the language barriers between people. Personally, I think this is going to lead to a meliorate world."

kowalskireatect1992.blogspot.com

Source: https://www.wired.com/2013/02/android-neural-network/

Share this post