'Thinking' computers get creative
Computers at a small St. Louis company busy themselves composing music, inventing consumer products and advising federal agencies.
They do this with little human input, said their inventor, Stephen Thaler, president and chief executive of Imagination Engines Inc. The machines are contemplative and deliberative, he said.
In other words, they think.
Thaler stands at the frontier of artificial intelligence, suggesting he has developed machines as smart as people. Other computer scientists disagree, contending that despite extraordinary advancements, it will take decades to achieve true artificial intelligence.
Virtually everyone thinks it will happen eventually.
Already there are computers with enough brute calculating force to solve problems once considered impossible. More recently computers have begun to help the operation of the human brain after it has been injured.
The next leap - the one a few scientists like Thaler believe they are making - is to develop computers that can conceive original ideas and can work with scientists much like human colleagues.
It's a far less imposing idea than the science fiction version of artificial intelligence, in which computers learn to outsmart their masters. But in a world where computers fly airplanes, handle financial transactions and have deeply insinuated themselves into every aspect of modern society, the limitations of the computer have become widely known.
For all their brainpower, computers can be dumb, misunderstanding simple instructions because they lack the intuition and flexibility of humans.
Computer scientists say it's taken longer than they ever thought, but there are now examples of what approaches a thinking machine.
At IBM's computational biology center in New York, a computer called Blue Gene is helping researchers study how human cells - including brain cells - operate. The machine is more than a supercomputer. It displays considerable intelligence and is even creating models to suggest how cells work, said Ajay Royyuru, senior manager at the center.
"So it's not giving me something that someone already knows, but in greater detail," said Royyuru. "It's actually giving me something completely different that no one had a clue about."
Among mainstream computer scientists, most regard Ray Kurzweil, a prolific inventor and author, as the most optimistic about advanced artificial intelligence.
Kurzweil predicts in his best-selling book, "The Singularity is Near," that in 2029 a computer will pass a test demonstrating human intelligence.
That test, proposed 50 years ago by computer pioneer Alan Turing, places a machine and a human behind a screen where they are questioned by a person. If, after extensive questioning, the person is unable to tell which is the machine, it will have passed the Turing test.
"You really have to capture the essence of human intelligence to pass a Turing test," said Kurzweil. "Through the right questions, you can unmask a computer: `Have you seen a movie lately? What did you think of the main character's motivation? Was he jealous?' Unless you're operating from a high level of intelligence, you just can't stay in a conversation like that."
Thaler, of Imagination Engines, is not interested in subjecting his computers to a Turing test.
"It's not that impressive," he said. "We could take a year out and build a system that appears to be like a human being. But you don't have to create something that's human. Machines aren't concerned with eating or keeping a job."
Thaler's machines are neural networks, which are electronic circuits configured by hardware and software to emulate human brains, learning through trial and error.
He introduces noise into one neural network to prod it into creating new ideas, while another neural network watches and provides opinions on which new ideas are useful and which aren't.
This is a machine equivalent of the human thought process, Thaler said.
"It gets smarter and smarter," he said.
Thaler's company uses the machines to do consultative work for the National Aeronautics and Space Administration, Department of Homeland Security and other government units, he said. His computers have also invented new products, such as an unusually designed toothbrush, and have written countless musical tunes.
While neural networks can be useful tools, they haven't achieved humanlike intelligence, countered Kristian Hammond, co-director of Northwestern University's intelligent-information computer lab. Hammond does expect that machines will achieve human intelligence.
"Of course we'll have at least human-level intelligence on a machine," Hammond said. "I've no doubt whatsoever."
He said this will likely come from countless small steps intended to make computers more useful, rather than from a focused drive to mimic the human brain.
"We look at a small amount of intelligence to do real problems," he said. "Fill in the gap. It could be if you fill in enough gaps, you'll build a whole system."
The most high-profile race to achieve human intelligence took place a decade or more ago at a chess board when Garry Kasparov took on IBM's Deep Blue.
"There was a lot of skepticism about that," Kurzweil said. "Kasparov played against computer chess programs in the early `90s and said they were pathetic," said Kurzweil. "In 1997, Kasparov lost to Deep Blue."
After winning the chess match, IBM scientists discussed mounting another grand challenge, but so far they've not devised one.
"We talked about playing Jeopardy," said Paul Horn, IBM's vice president and director of research. "Which is very difficult for a computer: to hear an answer, understand it, go through databases and find the single question that fits it.
"We decided if we worked hard, maybe the machine could beat a 4-year-old. It wouldn't be very impressive."
Deep Blue's successor, Blue Gene, is at work on studies of human brain cells.
Recently Blue Gene developed a model of what happens when a photon of light is absorbed by a protein found in a brain cell. Part of the reaction, which enables us to perceive light, was triggered by water.
That notion ran counter to everything scientists had expected, and it was revealed by the computer.
Follow-up experiments suggest the computer got it right.