Whatever happened to machines that think?

Whatever happened to machines that think?

• 
  23 April 2005
  
• 
  Justin Mullins
  

CLEVER computers are everywhere. From robotic lawnmowers to intelligent lighting, washing machines and even car engines that self-diagnose faults, there's a silicon brain in just about every modern device you can think of. But can you honestly call any machine intelligent in a meaningful sense of the word?

One rainy afternoon last February I decided to find out. I switched on the computer in my study, and logged on to www.intellibuddy.com, home to one of the leading artificial intelligences on the planet, to see what the state-of-the-art has to offer.

Intellibuddy: Hi, I'm Intellibuddy!

Nice to meet you.

You think I am cleverer than my editor.

You think I am cleverer than my editor.

You think I am cleverer than my editor...

So much for today's artificial intelligence. Intellibuddy is a version of one of the world's most successful chatbots, called ALICE (Artificial Linguistic Internet Computer Entity) and invented in 1995 by Richard Wallace, an independent researcher based in San Francisco. You can find versions of ALICE all over the web; the software is free. But whichever version you choose to chat to, the results are disappointingly similar. While some conversations have promising starts, all descend into the type of gibberish that only artificial intelligence can produce.

And it's not as if there hasn't been time to perfect the idea. The first chatbot appeared in the 1960s. Back then, the very idea of chatting to a computer astounded people. Today, a conversation with a computer is viewed more on the level of talking to your pet pooch - cute, but ultimately meaningless.

The problem with chatbots is a symptom of a deeper malaise in the field of artificial intelligence (AI). For years researchers have been promising to deliver technology that will make computers we can chat to like friends, robots that function as autonomous servants, and one day, for better or worse, even produce conscious machines. Yet we appear to be as far away as ever from any of these goals.

But that could soon change. In the next few months, after being patiently nurtured for 22 years, an artificial brain called Cyc (pronounced "psych") will be put online for the world to interact with. And it's only going to get cleverer. Opening Cyc up to the masses is expected to accelerate the rate at which it learns, giving it access to the combined knowledge of millions of people around the globe as it hoovers up new facts from web pages, webcams and data entered manually by anyone who wants to contribute.

Crucially, Cyc's creator says it has developed a human trait no other AI system has managed to imitate: common sense. "I believe we are heading towards a singularity and we will see it in less than 10 years," says Doug Lenat of Cycorp, the system's creator.

But not all AI researchers welcome such claims. To many, they only fuel the sort of hype that spawned what became known as the "AI winter" of the 1990s. This was a time that saw government funding for AI projects cut and hopes dashed by the cold reality that making computers intelligent in the way we humans perceive intelligence is just too hard. Many scientists working in areas that were once considered core AI now refuse even to be associated with the term. To them, the phrase "artificial intelligence" has been forever tainted by a previous generation of researchers who hyped the technology and the fabled singularity beyond reason. For them, the study of artificial intelligence is a relic of a bygone era that has been superseded by research with less ambitious, more focused goals. Arguably, this has already led to a limited form of AI appearing all around us. Elements of AI research are now used in everything from credit-scoring systems and automatic camera focus controllers, to number plate recognition in speed cameras and spaceship navigation.

AI is in many ways as old as computing itself. The purpose of building computers in the first place was, after all, to perform mathematical tasks such as code breaking that were too tough for humans to tackle. It was in 1950 that Alan Turing, the celebrated second world war code breaker, mathematician and arguably inventor of the first computer, formulated the test that would become the benchmark by which the intelligence of all computer programs would subsequently be measured (see "Turing's test"). Even in Turing's day, computers were beginning to outperform humans in certain specific tasks. And as early as 1948, John von Neumann, one of the fathers of the computer revolution, said: "You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that." It seemed just a matter of time before computers would outperform people in most mental tasks.

But many scientists and philosophers baulked at the idea. They claimed that there was something about being human that a computer could never match. At first the arguments centred on properties such as consciousness and self-awareness, but disagreement over what exactly these terms meant and how we could test for them prevented the debate from making any real progress. Others admitted computers could become intelligent but said they would never develop qualities such as compassion or wisdom which were uniquely human, the result of our emotional upbringing and experience. The definition of intelligence itself began to slip through the philosophers' fingers, and the disagreements continue today.

Most researchers would at least encompass in their definition of AI the goal of building a machine that behaves in ways that would be called intelligent if a human were responsible for that behaviour. Others would cast the definition even wider. Ant colonies and immune systems, they say, also behave intelligently in ways that are utterly non-human. But to get bogged down in the debate is to fall into the same trap that has plagued AI for decades. The Turing test is a reasonable yardstick. We will know an intelligent machine when we can talk to one without realising it is a machine, and programs of Intellibuddy's ilk clearly fall short of this requirement.

Intellibuddy is merely one of the latest in a long line of chatbots. In 1966 Joseph Weizenbaum, a computer scientist at the Massachusetts Institute of Technology, developed the first chatbot, named Eliza after Eliza Doolittle, the character in George Bernard Shaw's Pygmalion who is taught to imitate upper-class English speech. The program was designed to mimic a psychotherapist and conversed with its patient mainly through a simple rhetorical trick: it reworded the patient's statements as questions. For example:

...and so on. Eliza was programmed to spot key phrases in its interlocutor's sentences and plug them into preformed sentences of its own. It was hugely successful. The idea of talking to a computer astounded people, and there are even anecdotes of people developing emotional attachments to Eliza.

This early success contributed to a sense of optimism that the problems of AI could be overcome, much of it based on the idea that some kind of grand unified theory of mind would emerge that would offer up a scheme to create artificial intelligence on a platter. The late 1960s and early 1970s saw feverish speculation about the impact intelligent machines might have on the world and the advantages they would bring to whoever developed them. The computer HAL in Stanley Kubrick's classic 1968 movie 2001: A space odyssey summed up the visions being debated, and the fears they conjured up.

It was against this backdrop that Japan's Ministry of International Trade and Industry announced, in 1982, a programme called the Fifth Generation Computer Systems project to develop massively parallel computers that would take computing and AI to a new level. The scale of the project and its ambition were unprecedented and raised fears in the west that Japan would end up dominating the computer industry in the same way it had taken the lead in the electronics and automotive industries. If it developed truly intelligent machines there was no telling what Japan might be capable of.

An arms race of sorts ensued in which the US and Japan vied for supremacy. The US Department of Justice even waived monopoly laws so that a group of American corporations that included giants such as Kodak and Motorola could join forces to match the Japanese research effort. Between them they set up the Microelectronics and Computer Technology Corporation (MCC) and asked Doug Lenat, then a computer scientist at Stanford University, to lead it. The Defense Advanced Research Projects Agency (DARPA), the Pentagon's research arm, also began to take an interest, and injected huge amounts of funding into the field.

But progress was frustratingly slow, and as the hoped-for breakthrough failed to materialise splits appeared between groups taking different approaches. On one side were those who believed that the key to intelligence lay in symbolic reasoning, a mathematical approach in which ideas and concepts are represented by symbols such as words, phrases or sentences, which are then processed according to the rules of logic. Given enough information, the hope was that these symbolic reasoning systems would eventually become intelligent. This approach appealed to many researchers because it meant that general proofs might eventually be found that could simultaneously revolutionise several branches of AI, such as natural language processing and machine vision.

Ignominious ending

But by the early 1990s, it had become clear that the Japanese project was not leading to any great leap forward in AI. Things were no better in the US, as most of DARPA's projects failed to produce significant advances and the agency withdrew much of its support. The repeated failures of these so-called expert systems - computer programs which, given specialist knowledge described by a human, use logical inference to answer queries - caused widespread disillusionment with symbolic reasoning. The human brain, many argued, obviously worked in a different way. This led to a spurt of enthusiasm for new approaches such as artificial neural networks, which at a rudimentary level imitate the way neurons in the brain work, and genetic algorithms, which imitate genetic inheritance and fitness to evolve better solutions to a problem with every generation.

Neural nets got off to a promising start and they are now used in everything from computer games to DNA sequencing systems. It was hoped that with sufficient complexity they could demonstrate intelligent behaviour. But these hopes were dashed because, though neural networks have the ability to learn from their mistakes, all existing models failed to develop long-term memory.

In the AI winter that followed, research funds became difficult to come by and many researchers focused their attention on more specific problems, such as computer vision, speech recognition and automatic planning, which had more clearly definable goals that they hoped would be easier to achieve. The effect was to fragment AI into numerous sub-disciplines. AI as an all-encompassing field died a sudden and undignified death.

This fragmentation has had some benefits. It has allowed researchers to create a critical mass of work aimed at solving well-defined problems. Computer vision, for example, is now a discipline with its own journals and conferences. "There are people who spend their whole careers on this one problem and never consider the other pieces of the puzzle," says Tom Mitchell, an expert on AI at Carnegie Mellon University in Pittsburgh, Pennsylvania. The same is true of speech recognition, text analysis and robot control.

Simple common sense

But Lenat refused to give up. As an academic he had been working on building a database of common-sense knowledge, which he believed would be the key to cracking artificial intelligence. When funding for MCC dried up, he decided in 1994 to go it alone and spin off a company called Cycorp, based in Austin, Texas, to continue the development of the AI system he named Cyc.

The driving philosophy behind Cyc is that it should be able to recognise that in the phrase "the pen is in the box", the pen is a small writing implement, while in the sentence "the box is in the pen", the pen is a much larger corral. Lenat reels off examples where such common-sense distinctions can make all the difference. In speech recognition, the best way to distinguish between the two spoken sentences "I've hired seven people" and "I fired seven people", which can sound virtually identical, is to analyse the context of the conversation and what it means, rather than just look at the words in isolation. "Humans would never have this problem, but a computer program might easily be confused," says Lenat. "There's almost no area of AI that wouldn't benefit. It's the difference between usability and non-usability."

Cyc has been meticulously assembled to relate each fact to others within the database. It knows for example, that in the sentence "each American has a president" there is only one president, whereas in the sentence "each American has a mother" there are many millions of mothers. Cyc's knowledge is stored in the form of logical clauses that assert truths it has learned. It is based on the symbolic reasoning systems that failed to deliver in the mid-1990s.

But Lenat and his team have made huge leaps since then. One of the curious things about Cyc is that the more it knows, the more easily it can learn. The volume of common sense it has accumulated means it can begin to make sense of things for itself. And every new fact that makes sense is incorporated and cross-referenced into its database. Lenat maintains that the rate at which the system can learn depends on the amount of common sense it has about the world, and that today's AI systems perform so badly because they have close to none. "The rate at which they can learn is close to zero," he says.

One of Cyc's most impressive features is the quality of the deductions it can make about things it has never learned about directly. For example, it can tell whether two animals are related without having been programmed with the explicit relationship between each animal we know of. Instead, it contains assertions that describe the entire Linnaean system of taxonomy of plants and animals, and this allows it to determine the answer through logical reasoning.

Cyc now contains 3 million assertions. Impressive as that is, sheer numbers are not the point. "We are not trying to maximise the number of assertions," Lenat says. Rather, he wants to limit them to the bare minimum that will allow Cyc to collect data on its own. He says Cyc is getting close to achieving that number, and it is already advanced enough to query each input itself, asking the human operator to clarify exactly what is meant.

Sometime this year it will be let loose onto the web, allowing millions of people to contribute to its fund of knowledge by submitting questions to Cyc through a web page and correcting it if it gets the answers wrong. "We're very close to a system that will allow the average person to enter knowledge," Lenat says. He envisages Cyc eventually being connected to webcams and other sensors monitoring environments around the globe, building its knowledge of the world more or less by itself.

When Cyc goes live, users should expect to get answers to their questions only some of the time because it won't yet have the common sense to understand every question or have the knowledge to answer it. But with the critical mass looming, in three to five years users should expect to get an answer most of the time. Lenat has pledged to make access to Cyc freely available, allowing developers of other AI systems to tap into its fund of common sense to improve the performance of their own systems.

Lenat's optimism about Cyc is mirrored by a reawakening of interest in AI the world over. In Japan, Europe and the US, big, well-funded AI projects with lofty goals and grand visions for the future are once again gaining popularity. The renewed confidence stems from a new breed of systems that can deal with uncertainty - something humans have little trouble with, but which has till now brought computer programs grinding to a halt.

To cope with the uncertainty of the real world, the new programs employ statistical reasoning techniques: for example, a robot might measure the distance to a nearby wall, move and make another similar measurement. Is it seeing the same wall or a different one? At this stage it cannot tell, so it assigns a probability to each option, and then takes further measurements and assigns further probabilities. The trick, of course, is to ensure that this process converges on a solution - a map of the room. In practice these systems work most of the time, but the all too real fear is that the number of calculations could explode far beyond the robot's capabilities, leaving it hopelessly confused. Ways of coping with these situations are currently a hot topic of research.

Systems using the mathematical technique known as Bayesian inference have improved the performance of many AI programs to the point where they can be used in the real world. The despised Microsoft Office paper-clip assistant is based on Bayesian inference systems, as are pattern-recognition programs that can read text, or identify fingerprints and irises. Other approaches to AI have produced specialist programs that have played, and occasionally beaten, world champions in chess (New Scientist, 17 May 1997, p 13), draughts, and even poker (New Scientist, 20 December 2003, p 64).

But problems remain. Voice recognition only works usably in ideal conditions in which there is little or no background noise, and even then its accuracy is limited (New Scientist, 9 April, p 22). Chess programs are only capable of beating humans because the game itself can be reduced to a tree of possible moves, and given enough time computers are able to evaluate the outcome of each sequence of moves to find the one that is most likely to lead to checkmate. In the game of Go the board position is much harder for a computer to evaluate, and even modestly talented players can beat the most powerful computer programs. Robots have trouble negotiating obstacles that a five-year-old can dodge with ease. And if Intellibuddy is anything to go by, machines you can interact with, that understand what you are saying and react appropriately, are still some way off.

It's in your head

Where could the secret to intelligence lie? According to Mitchell, the human brain is the place to look. He has been using functional magnetic resonance imaging (fMRI) to see which parts of the brain become active when a person thinks about a specific object. He has found that when people are asked to imagine a tool such as a hammer or a building such as a house, the same areas of the brain are activated as when they are shown a picture of these objects. He has also found that the area activated for each object - hammer or house - differs by a discernable amount depending on the object. "You can train a program to look at the brain images and determine with 90 per cent accuracy whether that person is thinking about a tool or a building," he says. Such a program could, eventually, literally read your mind.

For the moment the research is confined to concrete nouns associated with physical objects. Next Mitchell plans to see how verbs make the brain light up, and beyond that whether the same areas light up when these nouns and verbs are incorporated into sentences, or whether there is additional activation elsewhere in the brain. Mitchell hopes that this approach will resolve the fundamental differences between the symbolic reasoning approach to AI, and biologically inspired approaches such as neural nets.

The clue that Mitchell thinks is significant is that the same part of the brain seems to be responsible for both reasoning and perception. So when thinking about a hammer the brain is acting like a symbolic reasoning system, and when recognising a hammer the brain is acting like a neural network. "Knowing this could provide some guidance," says Mitchell. Exactly how the designers of neural networks might use such findings is not yet clear. But Mitchell is convinced that this type of insight from functional brain imaging is set to have a huge impact on the field.

Of course if Lenat's prediction proves true, by the time Mitchell's work bares fruit, Cyc may well have reached the singularity. The history of AI suggests that is unlikely, but after decades of faltering starts and failed promises, things are beginning to change. Finally, machines might soon start to think for themselves.

A brief history of AI

1936 Alan Turing completes his paper "On computable numbers" which paves the way for artificial intelligence and modern computing

1942 Isaac Asimov sets out his three laws of robotics in the book I, Robot

1943 Warren McCulloch and Walter Pitts publish "A logical calculus of the ideas immanent in nervous activity" to describe neural networks that can learn

1950 Claude Shannon publishes an analysis of chess playing as a search process

1950 Alan Turing proposes the Turing test to decide whether a computer is exhibiting intelligent behaviour

1956 John McCarthy coins the phrase "artificial intelligence" at a conference at Dartmouth College, New Hampshire

1956 Demonstration of the first AI program, called Logic Theorist, created by Allen Newell, Cliff Shaw and Herbert Simon at the Carnegie Institute of Technology, now Carnegie Mellon University

1956 Stanislaw Ulam develops "Maniac I", the first chess program to beat a human player, at the Los Alamos National Laboratory

1965 Herbert Simon predicts that "by 1985 machines will be capable of doing any work a man can do"

1966 Joseph Weizenbaum, a computer scientist at the Massachusetts Institute of Technology, develops Eliza, the world's first chatbot

1969 Shakey, a robot built by the Stanford Research Institute in California, combines locomotion, perception and problem solving

1975 John Holland describes genetic algorithms in his book Adaptation in Natural and Artificial Systems

1979 A computer-controlled autonomous vehicle called the Stanford Cart, built by Hans Moravec at Stanford University, successfully negotiates a chair-filled room

1982 The Japanese Fifth Generation Computer project to develop massively parallel computers and a new artificial intelligence is born

Mid-1980s Neural networks become the new fashion in AI research

1992 Doug Lenat forms Cycorp to continue work on Cyc, an expert system that's learning common sense

1997 The Deep Blue chess program beats the then world chess champion, Garry Kasparov

1997 Microsoft's Office Assistant, part of Office 97, uses AI to offer customised help

1999 Remote Agent, an AI system, is given primary control of NASA's Deep Space 1 spacecraft for two days, 100 million kilometres from Earth

2001 The Global Hawk uncrewed aircraft uses an AI navigation system to guide it on a 13,000-kilometre journey from California to Australia

2004 In the DARPA Grand Challenge to build an intelligent vehicle that can navigate a 229-kilometre course in the Mojave desert, all the entrants fail to complete the course

2005 Cyc to go online

In his essay "Computing Machinery and Intelligence", published in the philosophical journal Mind in 1950, the British mathematician Alan Turing argued that it would one day be possible for machines to think like humans. But if so, how would we ever tell? Turing suggested that we could consider a machine to be intelligent if its responses were indistinguishable from a human's. This has become the standard test for machine intelligence.

In 1990, Hugh Loebner, a New York philanthropist, offered a $100,000 prize for the first computer to beat the test and a $2000 annual prize for the best of the rest. Past winners include Joseph Weintraub, three-time winner who developed The PC Therapist, and Richard Wallace, the man behind ALICE, who has also won three times.
Printed on Fri Apr 22 08:14:34 BST 2005

Directory: ld
ld -> Dsci 3870: Management Science
ld -> Primary holiday activities
ld -> Vitae weldon A. Lodwick Born: São Paulo, sp, Brasil – January 26, 1944 education
ld -> Vitae weldon A. Lodwick Born: São Paulo, sp, Brasil – January 26, 1944 education
ld -> Members of The War of 1812 Society in Virginia participate in a Plaque unveiling at the Old Donation Episcopal Church in Virginia Beach The program: Below the Master of Ceremonies, Dr Thomas Whetstone presides
ld -> Kristen Corbosiero, former daes post doc Ron McTaggart-Cowan, and McGill University Professor John Gyakum
ld -> State of Florida
ld -> -
ld -> The harmony society

Download 41.72 Kb.

Share with your friends: