Dueling Brainscapes In Artificial Intelligence

Two different approaches to building the cyberthinker of tomorrow

Meet Cog, the android wannabe--wannabe because it doesn't have legs yet. Those will come later. For now, it's still learning to coordinate its eye, head, and hand "muscles." You can almost believe it wants to grasp the hand of its proud father, Rodney A. Brooks, associate director of Massachusetts Institute of Technology's Artificial Intelligence Laboratory in Cambridge.

Cog (from cognitive) is the grand experiment in the latest approach to artificial intelligence (AI): letting a machine discover the world on its own, the way humans do, rather than cramming its memory with some preexisting computer model that describes the world from a human perspective. Call it the bottom-up approach.

The flip side of the AI coin can be found in Texas. Meet Cyc (from encyclopedia), the most ambitious version of the old-school, top-down system. Some $40 million has been spent on organizing Cyc's reasoning "engines" and stuffing its knowledge base with a half-million rules derived from 2 million common-sense facts. These are the things people soak up during childhood: Mothers are always older than their daughters. Birds have feathers. When people die, they stay dead.

Such facts may seem obvious, yet their omission regularly trips up AI programs, says Cyc's creator, Douglas B. Lenat, president of Cycorp Inc. in Austin and a consulting professor at Stanford University. But the size of Cyc's knowledge base is critical for another reason: "Learning occurs at the fringes of what you already know--so the more you know, the more you can learn." Starting from scratch, says Lenat, means Brooks "is in a race to create intelligence before the universe ends, since it took nature millions of years to get to us."

Brooks, a former student of Lenat's at Stanford, remains adamant. "My approach is the right way and will solve everything," he insists. "If Cog is successful, it can learn those common-sense things much faster than people can program them."

Mark Weiser, chief technologist at Xerox Corp.'s Palo Alto Research Center, likes Brooks's bottom-up tack. "Brooks is sailing west to discover America," he says. "Lenat is old paradigm--he hopes to find America in Portugal."

Lenat, however, figures he's already negotiating for Manhattan. Even though his vision for a knowledge-discovery system is far from complete, Cycorp is selling limited-purpose versions of Cyc. They've been snapped up by a half-dozen companies, including Glaxo Wellcome, Digital Equipment, IBM, and United Healthcare. Glaxo Wellcome and United Healthcare use their Cycs to manage huge online thesauruses of pharmaceutical and health-care terms that are a pain for people to deal with.

PHOTO SYNTHESIS. To show how Cyc's common-sense method can help find information that other software might miss, Cycorp has a database of captioned photos. Most database managers retrieve photos based on a precise word match in the caption. Not Cyc. Type in "strong and daring person," and Cyc pulls up a picture captioned "Man climbing mountain." Cyc knows that a man is a person, and that mountain climbing demands strength and is dangerous.

Cyc was hatched in 1984 at Microelectronics & Computer Technology Corp. (MCC), an Austin-based research consortium of high-tech companies. MCC funded a decade's work by a crew of linguists, philosophers, anthropologists, and engineers who spoon-fed data to Cyc. By the end of 10 years, Cyc was ready to earn its keep, so MCC spun off Cyc in 1994. Cycorp has turned a modest profit each year on annual revenues and research contracts of about $3 million.

The next step calls for Cyc to begin learning on its own by reading newspapers, books, and scientific journals. Then, in eight or nine years, Lenat figures Cyc will be smart enough for postgraduate work. It might help doctors make better diagnoses by checking medical records and presenting alternatives. Or it might help market researchers spot sales patterns missed by conventional data-mining programs. "This technology will pervade all software," declares Lenat. "There's almost no application that couldn't use common sense if it was cheap enough."

LATE START. By 2020, Lenat hopes Cyc will be ready to take charge of its own research lab. He expects Cyc to design unique experiments and uncover new knowledge.

MIT's Brooks has similar dreams for Cog's progeny, but the timetable is less certain because Cog got off to a later start. It was conceived just five years ago, after a Jan. 12, 1992, party that Brooks gave to celebrate the birthday of HAL, the AI system featured in 2001: A Space Odyssey. After brooding about the lack of anything close to HAL, Brooks decided he had to make a run at it.

Cog may be young, but it rests on mature technology. Brooks has been building self-learning robots for a dozen years, ever since the self-described bad boy of robotics decided that all of AI's past had to be chucked. He began building small, insect-like robots with simple brains connected to sensors. There's no central control, and walking is not preprogrammed. Yet the robots learn to walk through trial and error.

Take Genghis. It was an early six-legged robot about a foot long. Each leg has two force sensors and is programmed with a few simple behaviors, such as "when up, swing forward" and "when down, lift up." When Genghis is switched on, it lies on the floor, legs flailing. But once it manages a step or two, other programmed behaviors kick in to modify the movements of the legs based on feedback from the force sensors. Soon, Genghis is walking around, climbing over small objects, and using its antennas to steer clear of walls. For Brooks, the key to successful robots is designing the system so "the next action is always deducible from the current sensor readings."

These principles were used to build Ariel, a cheap crablike robot developed for the Navy that can locate and detonate mines on beaches. Another robot, Hermes, is a tiny version of Genghis that NASA wants to use to explore the surface of Mars. After the Mars Lander sets down, hordes of the robots would scurry. When one found something interesting, it would radio the Lander to come and collect a sample.

Brooks's success with insect robots reinforced his conviction that, just as human senses dictate what we learn about the world, a robot will develop human-like intelligence only if it has human-like senses. So Cog has two video-camera eyes and two ears. They're each linked to one of two dozen microprocessors, from Motorola Inc. and Texas Instruments Inc. Each chip controls a specific behavior or sensory system. "Last summer, we reached a critical point," says Brooks. "People started coming up to Cog and interacting with it like it was another being."

If all goes well as more behaviors are added, such as a sense of touch and then smell, Brooks knows what he wants the result to be: something like Lt. Commander Data, the supersmart android in Star Trek. How long might that take? Brooks throws up his arms. But maybe, around 2020, the dueling duo will mellow and give us Commander Cycog.

Before it's here, it's on the Bloomberg Terminal.