By Wilson da Silva
PROFESSOR Marvin Minsky is quick to the point. The bespectacled and slight man, one of the founders of artificial intelligence, says today’s computers already “think”, but they have a serious problem – they lack common sense.
“You and I know you can pull a string, but you can’t push it,” Minsky said in an interview. “Every child knows why, but I don’t think there’s a computer in the world that has a file on this. Find me a place in a book that explains that.
“Something that everybody knows won’t be written anywhere. Why bother? But you need to describe it in detail for a computer to understand it.”
Hence Minsky’s first law – before we can arrive at a machine, be it a robot or an intelligent car, that can interact with us the way we interact with ourselves, a mammoth catalogue of simple, everyday things will have to be amassed. Without such a common-sense knowledge base, computer brains will never be able to fathom the human world.
“You can say that machines are thinking, they’re just thinking dumb little thoughts,” Minsky said, sipping a cup of coffee. “They’re very precise and very fast, but they’re limited thoughts.
“What we haven’t been able to do is get them to do the things we think are easy. There’s a good reason for that – when you do something that seems easy, like recognising an object, that’s a big part of your brain that’s highly evolved and works so well that it doesn’t tell the rest of the brain how it’s doing it. It’s automatic, so it seems easy. But in fact, it’s very hard to do.”
Minsky is considered one of the fathers of artificial intelligence, or AI as proponents call it. He helped found the artificial intelligence laboratories at the Massachusetts Institute of Technology (MIT) in the United States, and through the 1950s and 1960s, built an international reputation with pioneering work at the labs.
He rallied a generation of bright young students to his side, and established some of the basic steps toward creating artificial intelligence. In the 1960s, his laboratory’s Saint program, the world’s first ‘expert system’, scored 96 percent on a calculus exam drawn from previous MIT first-year finals. A former president of the American Association of Artificial Intelligence, Minsky helped make MIT a world centre for robotics with his early work on computer vision and manipulation.
Minsky says 30 years of AI research has taught scientists that intelligence is not one simple, all-encompassing presence. It is most likely a co-operative association between a myriad of interacting ‘thought tasks’, controlled by a weak supervisory ‘thought task’ we would probably recognise as ‘consciousness’.
“If you’re holding a cup of coffee,” he said, holding up his white china cup away from its saucer. “You don’t want to have to think about whether the cup is tilting. What you do is, in the spinal column and the cerebellum, you set up little automatic sub-robots that keep measuring the pressure on your thumb and your finger, and if there’s more pressure on your thumb, it sends a message back to your wrist to rotate and keep the cup level. And that doesn’t bother the part of you that’s talking or doing other things.
Similarly, some believe the brain “is 300 or 400 rather complicated computers in a big network. You can’t do this sort of thinking by brute force. Thinking is too clever and tricky. But you can do it by 300 different approaches and somehow managing them to work together,” he said.
To mirror human intelligence, researchers would have to divide thinking into hundreds of smaller tasks. This would be supervised by a ‘controlling program’, what we would call ‘consciousness’ or ‘intelligence’ in the human brain.
“It’s the only good solution, working out ways of dividing the problem up, so you don’t have to have this one computer in the middle that knows everything. That would be too slow. And the ‘controlling program’ doesn’t have to be very smart. It just has to be there, and when there’s a conflict, figure out which of the two parties [or thought tasks] should be turned off.
“Consciousness is very weak. Consciously you don’t how you talk, or write, or see. People make a big fuss about consciousness, but it really doesn’t do much. You only use it when other things fail. Consciousness is a minor part of the thinking process, it will be very easy to put into machines.
“The hard part is, how do you retrieve the necessary facts to solve a problem. You know a million things – if the coffee isn’t sweet enough,” he said, motioning to his quickly emptying cup. “The answer is, ‘put some sugar in it’. Consciousness doesn’t give you the idea of the sugar, it just notes there was a problem, and a quick search solved it. Consciousness is very stupid. You don’t know how to talk, you only can talk.”
Minsky soon returned to the problem of imbuing machines with common sense, the sort of knowledge we humans take for granted.
“AI is waiting for these common-sense knowledge bases. I hope in about five years, some of them will become available. Otherwise, things will proceed very slowly. In my opinion, most applications are been held up by this.
“There are lots of applications designed for specific tasks that are proceeding. But the kind of machine that could talk to a person and help them with an everyday problem is still not on the horizon, because no machine knows all of the many meanings of ordinary spoken words.
“With a common-sense knowledge base, you could begin to solve everyday problems. You could make a robot that could clean the house and make the beds. You could make robot mechanics that could repair automobiles and fix torn clothing. There are so many things that people do [which can be done by machines], and in the end, everything that people do.”
He believes existing problems with making computers understand speech are surmountable: “There’s a certain amount of development in natural language in computers, understanding everyday speech instead of mathematical language. If.....you.....speak.....like.....this, there are already some low-cost machines that do a very good job. But if you speak continuously like this, there are some very big, slow machines that can do it, but they cost too much.”
“But that’s doomed until we get the common sense problems out of the way. With a common-sense knowledge base a machine could begin to solve everyday problems. If you can make a machine nearly as good as a person, then it should be easy to make it better than a person.
“The problem is, how do you get a machine to the level of a five-year-old child? If we get machines up to that then I don’t think there are any limits,” he said.
Early in the history of AI, some thought that the solution to giving computers intelligence was boosting their memory and processing power to dizzying levels, the so-called ‘brute force’ approach. Others thought they could build machines that worked on pure logic, and would be able to reason through the application of pure computing power to a problem. But this was found to not work well.
Today, AI scientists are concentrating on breaking thought into separate tasks.
“I think some of the best things happening now are programs that allow computers to learn. That would allow you to send robots to school to learn by themselves instead of being programmed. Send it to school or let it watch movies or read books – it’s worked for us.
“There are some learning programs already in use. In chemical factories they monitor production and make small changes to improve product flow, and learn by experience. That’s called an ‘adaptive system’. It learns rather quickly, and they’re used all over industry.
“I saw an elevator-control program that learned by experience which floors were the most popular at different hours of the day. So if an elevator isn’t in use – you know, in an old building usually an elevator ends up at the ground floor. But toward the end of the day, the best place for a smart elevator to be might be in the fifth floor of the building because the average time for the elevator to get to the 12th floor is shortest. It’s very hard to figure this sort of thing out, but some of the adaptive-learning machines do a fine job. An if conditions change in a few days, you don’t have to call up the factory to re-program the elevator.
Some big money is being spent on AI. At the International Joint Conference on Artificial Intelligence in August, 1,500 researchers from 31 countries met at the Sydney Convention Centre for six days. Topics ranged from “Parallel Non-Binary Planning in Polynomial Time” to “Emergent Frame Recognition and Its Use in Artificial Creatures”, detailed in 1,320-pages of conference notes bound like telephone books into two volumes.
Among the companies taking part were International Business Machines Corp and Sun Microsystems Ltd of the United States, Fujitsu Ltd and Hitachi Ltd of Japan, and Australia’s Broken Hill Pty Co Ltd.
Despite Minsky’s optimism, there are still problems for the world’s artificial intelligence experts. Delegates detailed studies into some of the remaining stumbling blocks – how to make computers learn from experience, apply their knowledge to new situations and be able to see and recognise the world around them.
Of them all, which is the next frontier for AI?
“Reasoning by analogy instead of by logic,” Minsky said. “The way humans solve things is, you have a problem, you can’t solve it but you say ‘well, this reminds me of another problem that I did solve’. Last time you had to change a light bulb, but last time there was a ladder. But there’s no ladder this time. ‘Last time I wanted to reach something high, though, I stood on a chair’ – it’s that sort of thing, people do this all the time.”
But it may take decades or centuries before computers can so totally match human intelligence. Until then, everyday appliances and machinery will become smarter, and a mix of AI and human interaction will become common.
“Remote control devices are something that I think could be done in a very short time. You could do jobs like fix a car in another country, if you could make robot hands that had pressure detectors on the fingertips to send back to you the same sensations you’d feel at the scene,” Minsky said.
Computer study has shed light on human intelligence, and suggests that “thinking” only accounts for 10 per cent of intelligent activity. The rest relies on learned facts and past experiences, with “thinking” linking the two and applying this
to each new situation.
“Psychology failed to tell us anything, and AI is replacing psychology as the science of the mind,” Minsky said. “Psychology just described what people did, but didn’t explain it. In AI, we have eight or 10 different ideas about how to make machines learn, psychology has only one or two and they weren’t very good.
“So now there’s a new field called cognitive psychology. It’s basically people explaining how old problems in psychology could be solved by computer-like processes. AI has revolutionised thinking about thinking. This new kind of thinking that relates computer processes to psychological processes is sweeping the world.”
Already there are about 500,000 robots working in industry worldwide. Intelligence would allow them to be placed in highly dangerous situations such as undersea mining or nuclear handling.
“I think we’ll see ‘intelligence’ being incorporated into everyday products...like vacuum cleaners and washing machines,” said Professor Michael McRobbie of the Australian National University’s Centre for Information Science Research. “Artificial intelligence will creep up on us as it becomes cheaper and cheaper to pump out silicon chips. You won’t even notice that machines around you can think.”