Google Uses AI to Make Search Smarter
(Corrects ninth paragraph to show that online advertising, and not just search, contributes almost 97 percent of Google's revenue.)
About 10 years ago, Peter Norvig helped develop software that temporarily operated NASA's Deep Space 1 spacecraft as it functioned without human help, more than 60 million miles from earth.
Norvig's current job as director of research at search giant Google (GOOG) keeps him focused on matters much closer to home. Norvig often shows up for work at the Mountain View (Calif.)-based company's headquarters known as Googleplex, wearing Hawaiian shirts and spending his days creating computer systems that help people search for answers to questions that aren't clearly defined. These can be as mundane as locating the best nearby pizza joint.
"It sounds not as dangerous as computers that take over the world, but it's something that helps with complexity and uncertainty," Norvig, 53, says in an interview.
The results of his work may be no less far-reaching than the exploration of Mars. It touches on how billions of people already use search, browse the Web, circulate e-mail, and translate documents and speech on personal computers and mobile devices.
In years to come, artificial intelligence (AI) systems might remind us of our appointments, drive our cars, and connect us with friends. "Imagine a very near future when you don't forget anything because the computer remembers," Google Chief Executive Officer Eric Schmidt said at a Sept. 7 conference in Berlin. "You are never lost. You are never lonely."
AI in action: Google Instant
For Google, which spent $2.84 billion last year on research and development, artificial intelligence is a big focus. "We think of that as something that permeates everything we do," Norvig says.
"Google's probably gotten the most number of people doing this," says James Allen, associate director at Florida Institute for Human and Machine Cognition, a nonprofit research firm affiliated with several universities. Google doesn't disclose the number of its staff working on AI-related projects, nor does it break out AI spending.
Artificial intelligence innovation could directly affect Google's revenue, which rose to $23.7 billion last year. Take Google Instant, a new search tool launched in early September that uses AI to guess what a person is searching for as they type keywords into the search box. Instant can save a user two to five seconds per search, according to the company.
The time savings should encourage people to turn to Google more frequently, ultimately lifting its advertising revenue, Susquehanna Financial Group analyst Marianne Wolk wrote in a Sept. 9 note. Online advertising contributes almost 97 percent of Google's revenue, according to Bloomberg data. "This innovation in user experience could improve Google's market share lead and at a minimum, force competitors to play catch-up," Wolk wrote. In August, Google completed almost 72 percent of all Web searches in the U.S., according to consulting firm Experian Hitwise. Rivals Yahoo! (YHOO) and Microsoft's (MSFT) Bing followed with 14.3 percent and 10 percent, respectively.
Artificial intelligence is pushing the use of Google on mobile devices, where it also dominates search. Google Instant, when it becomes available on mobile phones later this year, should increase the company's appeal to consumers on the go, driving up mobile ad prices, Wolk wrote.
exploiting dominance in user data
Speech recognition, another AI application, is already responsible for mobile search growth: In the U.S., one in four searches on a phone running Google's Android operating system is conducted via voice, the company says. "More people do searches because they can talk," Mike Cohen, manager for speech technology at Google, says in an interview.
Google's dominance in search and certain other applications gives it an edge in artificial intelligence. The company's smart software has billions of pieces of data to sift through and learn from. Among them are more than a billion search keywords typed into the search engine daily and 24 hours of video posted to YouTube.com each minute. Using the data, the software can make deductions about consumers' behavior, wants, and needs to provide them with more relevant answers.
Google "is most likely to be the sharp end of the stick" of artificial intelligence, says Geordie Rose, founder of D-Wave Systems, which creates chips for high-performance computers. D-Wave worked with Google on developing Google Goggles, an early example of so-called visual search enabled by artificial intelligence. Released in December, the mobile app lets an Android smartphone user snap a photo of, say, a menu item written in Spanish, and then get a translation in English. A user can also point the phone at a landmark such as San Francisco's Golden Gate Bridge to retrieve information about its history.
The app can't yet recognize food, plants, or animals, whose sizes, shapes, and postures vary dramatically. Google may need two to five additional years to teach the software to identify objects accurately, says Hartmut Neven, lead engineering manager for Goggles.
Smartphones that talk back
Eventually, Norvig would like the phone to be always on, digesting a user's surroundings—as well as text and speech inputs—to generate useful information. To that end, Google will this year unveil so-called "Conversation Mode" for its Translate software, which is already used 160 million times daily and which can translate among 57 languages. The software will make translations and speak sentences with a press of a button during conversations—say, between a tourist and a shoe clerk. Eventually, such translation should happen automatically, Norvig said.
The search engine will not only take input but also talk back. "Right now, most of the initiative is with the user, but we'd like the conversation to be more shared."
Battery life limitations would make it challenging to keep a phone working full-time. A further obstacle is the time frame in which Norvig and his team can work on projects. "When you are working on Mars missions, it's a 10-year cycle," Norvig says, "whereas at Google, the cycle is more like 10 weeks."
with Brian Womack