午夜福利1000集合

Data junkie

This article is听usually available only to听subscribers but is being made free to view thanks to sponsorship from听Ocado

听

Take a mind-bogglingly huge quantity of data and apply some simple statistics: that鈥檚 Peter Norvig鈥榮 strategy for changing the way we live

PETER NORVIG doesn鈥檛 suffer fools gladly. Nor bureaucracy. Nor journalists who fail to check their facts. This much is clear from browsing . Among papers on artificial intelligence nestle idiosyncratic essays such as the scathing , plus , brought to you by a Bern patent office that has somehow travelled through time to purchase a modern human resources handbook.

Norvig is also a world leader in computer science, who prior to joining Google a decade ago headed what is now the at NASA鈥檚 Ames Research Center, developing software to allow robotic spacecraft to operate autonomously. So as I drive down the San Francisco peninsula to our meeting at Google HQ, a stone鈥檚 throw from Ames in Mountain View, it is with both anticipation and trepidation. At least he has a sense of humour, I tell myself.

That Norvig has a compelling vision becomes clear when he launches into an overview of Google鈥檚 research. His big idea is that if you can amass sufficient data, a few relatively simple statistical algorithms are enough to solve some of the most vexing problems in machine learning, such as automated language translation. 鈥淚n the past most people said, 鈥榃e鈥檒l stop when we hit the bounds of what can fit into the memory of one computer鈥�,鈥� Norvig says. 鈥淲hereas we said, 鈥榃e鈥檒l stop when we fill up a data centre鈥�.鈥�

The business world is full of analysts touting the power of 鈥渂ig data鈥�. Much of it is arm waving, but at is 鈥渢o organize the world鈥檚 information鈥�, Norvig is confident that he has found the real deal. This, he argues, is why Google can instantaneously translate web pages between dozens of languages; why it is leading the way in visual search; and why one day you may talk to machines in colloquial, heavily accented speech and have them understand you, first time.

are the most mature product of this approach. 鈥淚n the past, people had thought of this as being a linguistics problem,鈥� says Norvig, which meant designing software that could replicate a human translator鈥檚 understanding of the languages involved, including their grammatical rules.

Instead, Norvig鈥檚 team has compiled texts that have already been translated, and applies statistical techniques to train the system to learn translations of unfamiliar words and how they are used in context. 鈥淓ssentially we are just building this big model of probabilities,鈥� Norvig says.

He explains the principle by recalling a visit to Berlin: 鈥淭here鈥檚 a brochure in my hotel, and the left side is in English and the right side is in German. If anybody asked how to say, 鈥楾o dial the operator press zero,鈥� I would know how to translate that. I don鈥檛 have to understand anything.鈥� Through this approach writ large, Google can now offer acceptably meaningful translations between more than 50 languages.

The same concept of applying simple statistical rules to vast quantities of data underpins the two other main thrusts of Google鈥檚 research: in visual search and speech recognition. Run through enough pictures tagged as the Eiffel Tower, Norvig says, and you don鈥檛 need massively complex algorithms for computers to learn that the tower is the pointy thing.

This will allow more powerful ways of linking real-world queries to Google鈥檚 familiar search capabilities, using mobile devices. 鈥淚f you鈥檙e in a store and you want information on a product you could type in its name, but just taking a picture seems easier,鈥� says Norvig. Through , smartphone users can already search for information on some objects, such as landmarks or bottles of wine, by snapping them with the device鈥檚 camera.

Speech recognition is a tougher nut to crack. Because of the almost infinite variety in accent, timbre and other characteristics of human speech, the conventional approach has been to train systems to learn individual users鈥� vocal tics. That鈥檚 fine if you want a personal dictation machine, but useless for allowing computers to interpret anything that鈥檚 said to them, by anyone.

Norvig is convinced that speech recognition will fall to the 鈥渂ig data, simple algorithms鈥� approach. The problem is finding enough data, as the spoken word is not represented online as comprehensively as text and images. As we discuss this issue, Norvig makes a revealing admission about the launch of , which among other things transcribes phone messages and sends them to your email inbox: 鈥淥ne of the reasons we had this phone service is that we wanted to capture lots of interactions; hear different accents and different voices saying different things.鈥�

No human is listening to your messages. Norvig simply means that computers are using the data to improve their ability to transcribe speech. But it鈥檚 this type of routine processing of personal information that makes some people uneasy about Google鈥檚 reach into our lives 鈥� and helps explain the company鈥檚 clashes with campaigners for online privacy.

Google Voice also illustrates a crucial difference in culture between Google and Norvig鈥檚 last employer. In the early days of the service, some commentators 鈥� New Scientist included 鈥� had fun at Google鈥檚 expense, noting the gibberish in some of the transcribed emails. Though things have improved since then, the transcriptions are still far from perfect, as Norvig admits. The company is relaxed about letting rudimentary versions of its products loose on customers, he says, because that helps gather more data and solicits feedback on what people want to see improved.

Such an approach could never be tolerated by NASA, where a single glitch can mean the loss of a mission. 鈥淵ou can鈥檛 send out a repairman,鈥� Norvig observes. So despite excellent results with the experimental Deep Space 1 mission to comet Borrelly, which tested a range of advanced technologies, NASA decided not to use on-board control software designed by Norvig鈥檚 team to give the Spirit and Opportunity autonomous control over their movements. Norvig well understands NASA鈥檚 caution, not least because of his experience investigating the failure of the Mars Climate Orbiter: the probe burned up in the planet鈥檚 atmosphere in 1999 because of confusion between imperial and metric units in the navigation software.

The inquiry also triggered in Norvig a visceral reaction to 鈥渂ullshit鈥� PowerPoint presentations, as officials addressed the investigating panel with slick but unilluminating slide shows. A few months later, Norvig created the 鈥溾��, a satire on the propensity for presentation software to blunt rhetoric and obscure meaning 鈥� especially if the presenter follows suggestions made by PowerPoint鈥檚 autocontent wizard.

Norvig realised it was time to leave NASA. 鈥淭he excitement had begun to wear off, and the dulling effect of bureaucracy never wears off,鈥� he says. Looking at the relaxed figure before me, it鈥檚 clear Norvig has found his spiritual home. 鈥淧lease don鈥檛 offer me a job. I already have the best job in the world,鈥� his website tells headhunters.

鈥淧lease don鈥檛 offer me a job. I already have the best job in the world鈥�

Does he worry about all those bright young things at Google who may want that job for themselves? Norvig says he wouldn鈥檛 mind if someone else took his position 鈥� just as long as he鈥檚 still allowed to play with Google鈥檚 oodles of data. 鈥淭he best part is being here,鈥� he says.