By In 1950 English mathematician Alan Turingââ,¬"famous for his role in cracking the German Enigma cipher during World War IIââ,¬"proposed that any computer able to trick people into believing it was human must be intelligent. Since then, people have been fooled, leading artificial intelligence theorists to ask again, "What is intelligence?" The answer for now is, we don't know. But that hasn't stopped researchers at the Massachusetts Institute of Technology and Xerox PARC from designing systems that act independently of human beings, learn from experience, and make decisions.
We've already grown accustomed to delegating some of our thinking to machines, through search engines, speech recognition, and natural language processors. AI's long-standing appeal dwells in the romance of mingling the creative problem-solving methods of human thought with the presumably flawless logic of computer circuits -- in a sense, building intellectual workhorses.
Thomas Poggio, Whitaker professor at MIT's Department of Brain Sciences, has helped produce quasi-intelligent technologies we now take for granted, including search engines and financial forecasting applications. But the work he's doing now has even deeper ramifications.
One project: bioinformatics.
Analysing a string of 7,100
human genes, AI-based bioinformatic chips detect certain cancers and prescribe appropriate medication, learning from each case encountered. "There is a lot of interest in the applications of this for the short term," Poggio says. "If you have high blood pressure, a doctor will try one of five different families of drugs. If the first one doesn't work, he'll move on to the second." Through bioinformatics, computers can analyse the gene expression of the disease, even if human operators don't understand the nature of the
genes, and "they can find out immediately which drug is
the right one."
Poggio has spent most of his career designing machine vision systems. "What we're trying to do is have a system that can be trained, rather than preprogrammed, to find objects in a photograph," he says. "We show it images...and the system learns to find the people in the images."
What's in a face?
It's much harder for computers to recognise the general
shape of the human face -- something humans do naturally -- than to recognise individuals by measuring their specific features (aka biometrics). Future applications of this work include Defense Advanced Research Projects Agency networks that can track an individual's movements worldwide with video surveillance cameras, recognising troublemakersââ,¬"or anyoneââ,¬"when their faces appear onscreen.
Like recognising a picture, navigating the physical world requires intelligence. Mark Yim, a senior Xerox PARC researcher, is tackling this problem with a new kind of robot to be used by DARPA for search and rescue missions. This robot is composed of independent modules linked together, with each 5-centimetre module sporting its own PowerPC processor. Every segment acts autonomously.
"By themselves, they don't do very much," Yim says, "but when you put a lot of them together, you can start to build complex machines."
Each self-controlled module communicates with its neighbours to decide how it has to move. While testing now involves human-directed teleoperation, Yim expects his robot eventually to go its own way. "At the highest level," he says, "the person would say, 'There's a rubble pile. Go find someone.' And the robot would figure out how to do it."
2%
4%
