The United States will be "betting the future of its entire information processing industry" if it does not counter the new Japanese 10-year crash plan to build a new generation of "smart" computers, computer scientist Edward Feigenbaum said at a computer conference here today.
Feigenbaum, a Stanford University professor and a leading figure in studies of artificial intelligence, or the development of smart computers, said the U.S. response to the Japanese program has so far been "discoordinated" and "not very promising" even though the stakes are the dominance of an entire industry in the future.
The new Japanese program was outlined in some detail by its director, Kazuhiro Fuchi, who also addressed the New York Academy of Sciences computer conference. The project is intended to build what is known as the "fifth generation" of computers, those that are many times faster than current ones and also are built chiefly as knowledge-manipulating machines rather than as number-manipulators.
Such a powerful knowledge-based computer could be the basis for a new generation of commercial products and services, such as "expert systems" like those that now give professional advice to doctors on diagnosis in some limited speciality areas, and to chemists analysing the molecular structure of substances.
The Japanese will spend at least $200 million in government funds on the new project, Fuchi said. Feigenbaum added that with additional grants from Japanese industry, the project will get about $850 million over the next 10 years. In each of the first four computer generations--those based on the vacuum tube, the transistor, the early integrated circuit and the current densely packed computer chips--computing power was advanced by hundreds to thousands of times.
The fifth generation computers will now concentrate on arrays of hundreds of chips, each computing independently, to multiply the speed of the machine. The machines will gear their fundamental way of operating not to numerical computations but to knowledge processing that uses language and other symbols.
Feigenbaum said that right now, the United States leads the world in basic research on intelligent machines. But he said that historically Americans have been extremely slow to take industrial advantage of their research power. Now, he says, the Japanese could outmaneuver the Americans unless something is done to compete.
Current American attempts to move toward the fifth generation machines are scattered among 20 universities and companies, he said. And he added that even though some of the very best research is being done at corporations, those companies are not taking full advantage of the research.
Also at the conference, a number of computer scientists assessed the state of the art of machine intelligence. They agreed that computers have now been programmed to do tasks that were once thought to require much intelligence. But at the same time, the scientists have discovered that it is the things humans take for granted about behavior that are not likely to be achieved soon by computers.
As Michael Dertouzos of MIT put it, "We can make computers that play world-class chess . . . but we can't make one that can walk into a room and find the chessboard."
The difficulty is that what appears to be complex behavior to humans--mathematics, chess--can be carried out using precise calculations. Computers can handle these. But what computers have not been able to handle are tasks involving pattern recognition and ambiguity, such as viewing a scene in a room and identifying the objects in it.
Machine translation efforts failed because human language is filled with ambiguity, as in the often-cited example of a sentence that a Harvard computer failed to translate: "Time flies like an arrow." The computer could not tell, for example, whether "time flies" might be a variety of fly, like blue-bottle flies, and if the sentence was about the shape of insects.