Most Americans may be oblivious to artificial intelligence, said researcher Bruce Buchanan, but soon it will be hard to avoid.
The AI community has made stunning–and inaccurate–predictions before. In the 1970s and ’80s, for instance, researchers boasted that by now computer systems would understand spoken language and carry on fluid conversations with us. Public interest in AI flared briefly, but when scientists’ predictions failed to materialize, it waned considerably.
Yet that experience led to two realizations in the industry, said Buchanan, a University of Pittsburgh professor and president of the American Association for Artificial Intelligence, which met here this month along with researchers from the Innovative Applications of Artificial Intelligence.
The first was that computing power had to be boosted considerably before most AI theories could be put into practice. The other was to deliver before talking a big game.
AI now seems poised to do just that.
Coming into its own within the last 10 years, AI research has drawn enormous benefit from parallel booms in networking, wireless communication and the Internet.
More powerful computers not only make individual AI functions possible, but also allow them to be combined. And, increasingly, they are being combined to the benefit of commercial products, such as Internet search engines that recognize patterns in data (an earlier spinoff from AI technology) and then learn which types of search results their users typically look at (an AI spinoff now beginning to hit the market).
And, perhaps more promising, researchers have found programming techniques that work in many different AI subfields, a development likely to ensure dramatic crossover breakthroughs that will allow AI to leap ever more deeply into our everyday lives. Think of refrigerators that make up shopping lists as they grow empty, electronic office assistants that shuffle our schedules–even cars that drive themselves.
But for the time being, AI people are downplaying their industry’s promise, even as they quietly slip it into our lives.
Hence the AI-powered Tip Wizard in newer versions of Microsoft Word, the script recognition in Apple’s Newton, and the AI programming in Tiger Electronics’ Furby dolls. Do we know that AI makes those things possible? Probably not, Buchanan said. And that’s the point.
Just because AI has been unobtrusive doesn’t mean the science behind it isn’t remarkable. Over the last decade, the field has enjoyed broad advances on a number of fronts–the kind of success that sparks enthusiastic debates among researchers on the comparative merits of programming through symbolic logic (as in flowcharts) or neural networks (which react quickly to outside stimuli).
Those approaches are, in turn, used to solve problems within AI’s subfields. Expert systems run through lists of likelihoods to make assumptions. Pattern recognition compares probabilities to evaluate large amounts of data, such as photographs or speech. Machine learning devises rules based on observations. Significant understanding has grown in each area–so much, in fact, that AI purists no longer consider them signs of intelligence.
“As soon as something works, it’s no longer considered AI,” said Chuck Thorpe, a principal research assistant at Carnegie Mellon University’s Robotics Institute. “As soon as something is understood, it’s spun off and becomes its own field.”
On the other hand, subfields complement each other nicely and can be combined with surprising results. Ron J. Brachman, a research vice president for AT&T Labs, said several groups are working on a help program–an expert system–with speech recognition. That way, you could call a support line, talk instantly with a computer that walks you through your problem, and solve simple issues in minutes.
“That technology is in a sense this close,” Brachman said. “It’s working in the labs, and you can go a surprisingly long way with what an AI purist would think of as shallow-knowledge processing.”
Robots, which perhaps have the most to gain from combining AI tools, have been able to get a lot of mileage from the approach.
Cerebus, a bare-bones robot brought to the meetings here by Ian Horswill and a team of graduate students from Northwestern University, can respond to basic questions, avoid bumping into people in crowded hallways, and follow someone at a distance when told to do so. Despite obvious rough edges, Cerebus has a lot on his mind, Horswill said. The robot still considers information from surroundings, plans routes in which to travel and knows enough to respond out loud to typewritten queries. And really, he said, Cerebus is a rough draft of robots to come.
“I think service robots, at least the underlying technology, will be commonplace within about five years,” said Alan Schultz, head of the Naval Research Laboratory’s Intelligent Systems Section.
Future robots likely will find careers in military reconnaissance, hauling office supplies or other equipment, and working in coal mines. “We talk about the three D’s,” Schultz said: “Dull, dirty and dangerous.”
Most of us won’t be saying hello to R2-D2 by the office water cooler, but if the computer industry has anything to say about it, we are likely to be talking to our computers soon, said Eric Horvitz, a senior researcher at Microsoft.
A user might talk to the computer “to clarify understanding about a project,” Horvitz explained. “Just as you would with a colleague.”
At Microsoft’s headquarters in Redmond, Wash., Horvitz works with a group of researchers to create useful, easy-to-talk-to office assistants.
One of their inventions tracks personal calendars and plans made by e-mail, pointing out scheduling conflicts as they occur. Another creation prioritizes e-mail messages so “Question from boss” shows up in a list before “Funny joke.”
Adding voice recognition to those applications would improve them greatly, Horvitz said, and the ultimate goal is to get computers to converse like people, figuring things out by asking questions until understanding is gained. Another step will be getting systems to function in the uncertainty of changing situations.
From talking to other researchers, one gets the idea such innovations are no longer impossible dreams. Buchanan said it would not be complicated to install sensors in a house that allow a home computer to figure out what room you are in. It could then turn off lights in empty rooms to save electricity, for instance, though Buchanan said that would by no means be the limit of how AI-empowered electronics could be useful in the home.
With a little extra AI, he said, the dishwasher could decide how dirty the dishes are and therefore decide how long to run the scrub cycle. The refrigerator could communicate with the date book on your personal computer and warn that you’ll need to buy more hot dogs before the barbecue you scheduled for Saturday. Or it could communicate with your car and start defrosting the chicken you planned for dinner as you are leaving the office.
If it sounds as though a big element in the presumptive Artificial Intelligence Revolution involves communication between devices, researchers generally agree that’s correct. But as computer power increases and computer size shrinks, it also seems likely that AI programming will grow yet more complex, in turn allowing it to become still more adaptive and less noticeable.
But even so, building better machines–even seamlessly interconnected machines–is only half of the AI community’s goal.
“There’s sort of two Holy Grails,” said Thorpe at Carnegie Mellon. “The engineering Holy Grail is to make things natural and unobtrusive. The scientific Holy Grail is to understand how people think.”
