Caltech Biophysicist Dreams of Ways for Computers to Think : Science: John J. Hopfield designs neural networks, vast matrices of interlinked computers modeled after the human brain.
In his years of research, Caltech biophysicist John J. Hopfield has met a number of people who think like computers. But he has never met a computer that thinks like a human.
The dizzying potential of such a machine has long intrigued him, however. A pioneer in the study of “thinking computers,” Hopfield is fascinated by whether digital machines can be taught to process information in a manner similar to the human nervous system.
Earlier this month, the Caltech professor gave an open lecture at Claremont’s Harvey Mudd College that probed the question “Can Computers Think?”
In contrast to traditional computers, which process information sequentially, bit by bit, Hopfield designs neural networks, vast matrices of interlinked computers, modeled after the human brain, that can communicate with each other and trade data all at once. Although the field is a fledgling one, Hopfield and others have laid the groundwork for radically new computers that could have enormous impact on technology.
The field appears tailor-made for Hopfield, who strides comfortably through the realms of neurobiology, computer science, physics and mathematics and in 1983 won a MacArthur Prize.
Hopfield, “has been a very major figure in the field for a long time,” said James A. Anderson, a cognitive scientist at Brown University. “He produced a very simple model that’s quite elegant.”
Winning the school’s prestigious Wright Award, presented each year for leading research in interdisciplinary fields, also puts Hopfield in eminent company. Last year, the $20,000 award went to British scientist Francis H. Crick, who shared the Nobel Prize with James D. Watson in 1962 for his discovery of the molecular structure of DNA.
Formerly, neural network research fell through the cracks of biology and computer science because it fit squarely in neither. But today, as scientists seek to design ever-smarter and more powerful computers, Hopfield is riding a growing wave of interest.
“There was a period in the 1970s when no one was paying much attention, except the hard-core, such as Hopfield. He is responsible for the resurgence,” Anderson said.
Work in Gray Areas
Hopfield, a tall, lanky man of 56 who explains his theories in clear, graceful language, attributes his success to his exploration of the gray areas where no scientist has gone before.
So do others.
“Had he not asked questions in areas that were in the boundaries between disciplines, he wouldn’t have made such a creative, profound contribution,” said Stavros Busenberg, a mathematics professor at Harvey Mudd.
All too often, however, traditional research methods have encouraged specialization at the expense of gaining fresh ideas from cross-pollination, Hopfield said.
“There’s lots of rewards for someone who’s at the center of his or her discipline. At the fringes, you don’t have a built-in rooting section, and you wonder if anyone’s really listening,” Hopfield said.
He has personal experience with this phenomenon.
Two Disciplines
“When I was in the Princeton physics department, they told me, ‘We’re sure what you’re doing on the outskirts of biology is very interesting, but please don’t tell us too many details.’ ”
But that is all behind him. For the last decade, Hopfield has held a joint appointment in the departments of chemistry and biology at Caltech, a fringe-dweller’s paradise.
He teaches only one class, in neural networks. It is a task he relishes--because he can discuss ideas with students--and dreads, because it forces him to anticipate students’ questions and answer them first.
But there are payoffs. In 1984, he said, on a flight home from an East Coast seminar on neural networks, fear of showing up unprepared for class prodded the Pasadena professor to construct a mathematical model that had long eluded him.
“A lot of the interesting things I’ve done in my life have been the result of having to give a lecture very soon. It forces you to think deeply.”
Hopfield has been thinking deeply about the connection between computers and the human brain since 1969, when a scientist friend convinced him that he should combine his physicist’s background with his interest in neurobiology.
‘Dim Ideas’ Linked
For several years, Hopfield signed up for seminars and collected information, slowly coming up with “the dim ideas that there was something collective about the way memory worked.”
Then, during three months in 1980, Hopfield made a connection between memory and neurobiology. But he says it took him several more years to come up with the idea of redesigning digital computers to reflect the architecture of the human brain.
In one experiment, Hopfield and his associate, David Tank at Bell Labs in Holmdale, N.J., used a simulated neural network to solve the classic “traveling salesman” problem, which involves finding the most efficient route for a salesman to travel among a set of cities.
Hopfield and Tank’s solution was to design a circuit board with a series of switches adjusted to represent the distances between cities. The computers worked in tandem, settling into a pattern that indicated the shortest route 1,000 times faster than a digital computer, which would have to measure each route one at a time.
Less Than 100% Sure
Like the human brain, the neural networks don’t give the right answer 100% of the time, Hopfield said. But in the traveling salesman test, they gave the best answer 50% of the time and one of the two best answers 90% of the time.
Of course, digital computers will always provide fast, efficient number-crunching, a skill at which humans don’t excel, Hopfield said. But they are just plain dumb when it comes to tasks that humans do effortlessly, such as recognizing people or voices.
Today, a new breed of computer pioneered by Hopfield and others is learning through association, using building blocks of data and experience to solve problems.
Hopfield is working with a graduate student on a neural net system that will be able to hear words and distinguish them from garbled letters. When it hears a bona fide word, a bulb will light up.
Such systems have many applications. Since 1986, hundreds of little companies have sprung up to try to build computers that apply the principles first articulated by Hopfield and a handful of others.
And a large group, the Pentagon, sees defense applications. Last year, it announced development of its own $33-million neural network system.
But for now, Hopfield said, thinking computers remain in the science fiction future.
Paraphrasing author and mythologist Joseph Campbell, Hopfield added: “At least for the near future, the computer is going to continue to seem like an Old Testament God, with lots of rules and little mercy.”
