MIT: Making Computers Smarter Than Humans

Professor Tomaso Poggio works with eyes. He does not treat them; he makes them--from scratch.

Poggio and his student proteges at MIT's artificial intelligence (A.I.) lab are in the process of building a machine which may someday be able to perform all the functions of a human eye: detecting the position, motion, shape and texture of almost any object.

Artificial "eyes" have been built before. "The most successful have only been able to deal with a constrained world: blocks and cubes painted a uniform color under controlled lighting," says Poggio. When complete, Poggio's invention will be able to recognize a human face.

Poggio's project is part of a decade-long effort at MIT to develop computers or robotic devices which can duplicate all the major capabilities of the human body, including sight, hearing and speech comprehension, touch, manipulative ability, locomotion, and reasoning power.

MIT, along with Stanford and Carnegie-Mellon, is at the forefront of artificial intelligence research, scientists there claim.

About 15 faculty, 85 students, and 60 staff members work at the lab. They research in small groups, each typically composed of a single faculty supervisor and several students and technicians. Each group devotes itself to mimicking a particular bodily function.

A researcher's typical "workday" consists mostly of writing programs, attending seminars, and just sitting and thinking about knotty problems, says Professor Carl E. Hewitt, a scientist at the lab. Hewitt says scientists at the lab spend surprisingly little time doing "lab work." "The lab is the computer," says Poggio.

Competition for the student research posts is intense; last year only 10 students were selected from a pool of 300 applicants. The great majority of student researchers are from MIT, but one, Oren W. Etzioni '86, is a Harvard undergraduate.

MIT graduate students work with Dr. John M. Hollerbach on perfecting a four-fingered robot hand, which he invented several years ago in collaboration with the University of Utah. "It was a 'joint' project," he quips.

In duplicating the human hand, Hollerbach says, all the myriad capabilities of flesh must be reproduced. Not only the ability to grasp, but also the ability to sense mass, configuration, texture, curvature, and temperature of various objects must be instilled in the machine. Hollerbach says that, if successful, his invention will be able to do all these and more, and will, in addition, have the same degree of mobility as a real hand

When perfected, the hand will be connected to--what else--a robotic arm, and this is where Professor Tomas Lozano-Perez and his fellow researchers come in.

Lozano-Perez has come up with an algorithm which allows a robot arm to carry out a pre-programmed series of tasks while avoiding obstacles which are placed in its way. He now is trying to come up with methods by which robots can "fiddle" with machine parts in order to make them fit.

Currently, says Lozano-Perez, industrial robots are no more than sophisticated "tape recorders." They are programmed by experts who lead their mechanical arms through, the motions of a particular task. The robots store this sequence of movements in their memories and can later repeat it an unlimited number of times.

Although such blind repetition may be sufficient for commercial applications, says Lozano-Perez, it does not exploit the full potential of the robot. He hopes to construct robots that need only general instructions, instead of the highly specified directions that present-day robots require. His creations would then by themselves identify needed parts, convey them to pre-ordained locations, and work them around--as a human might--until they fit properly.

The crowing glory of a successful imitation of the human body would certainly be duplication of the thought processes of the human brain. Professor Carl E. Hewitt is exploring possibilities in this area.

Modern computers are usually "the exact antithesis" of the human bram, says Etzioni, who works in conjunction with Hewitt. While the brain works relatively slowly yet can do several things simultaneously, the computer runs quickly but can handle only one thing at a time.

Hewitt says he wants to build a computer network which works like "a human organization." In his network, computers would be in communication with each other but would retain the ability to act independently.

Hewitt's goal can be more clearly understood by using the analogy of a court of law, says Etzioni. In making decisions, several computers would contribute evidence supporting one point of view, while others would offer conflicting arguments. Through their intercommunication, the computers would together be able to arrive at a "verdict" which they could not have obtained separately.

Artificial intelligence researchers have set their sights high, and are optimistic about the chances of major breakthroughs in coming years. Most scientists at the labs feel that there is no fundamental limit to how closely machines can be made to resemble humans.

"I don't think there is anything that computers will not be able to do at some point," Poggio says.

Lozano-Perez says that one of the fastest-developing areas of A.I. at the moment is artificial hearing.

Computers which can hear and interpret single-word commands have been around for several years, he says, but only now are computers being adapted to comprehend continuous speech. Within the foreseeable future, therefore, computers may be able to verbally interact with humans.

Poggio says he predicts that within 10 years, "you will be able to buy, for the price of a car, a home robot that will clean the floor, make the beds, cook for you, etc."

Other scientists are more restrained in their estimates. Hollerbach says robots like the one Poggio described are probably a long way off. For the present, he calls attempting to embody all artificial intelligence research in a single robot "a waste of time," because while the research remains primarily in the theoretical and experimental stages, "the whole will be less than the sum of its parts."

"The kind of things you read about in science fiction are far away, and the closer we get, the more problems we run into," says Jim V. Mahony, an MIT graduate student who works with Poggio.