AI

A machine that behaves like a human. Technology that is derived from dreams. The modeling of human intelligence. An entity that understands it will die. An enhancement to natural intelligence. Getting machines to do what humans do well. It has no meaning for the average user. Systems that use some form of reasoning...

These were some of the answers received when I put the question, "How would you define artificial intelligence?" to a number of Japanese, or Japanese-resident, academics and business people. The most prescient comment came from Ryohei Orihara of Toshiba's Knowledge Media Research Laboratory, who said: "Getting everybody to agree on a definition will be the hard problem."

This selection of academics, entrepreneurs, and salarypeople all came from a diverse group whose work can be linked to the nebulous concept of artificial intelligence (AI), as popularized by the recent Steven Spielberg film of the same name. Norihiro Hagita of the Japanese Society for Artificial Intelligence (JSAI) explained that the organization's "What's AI?" Web site had seen a 700 percent increase in traffic following the release of the film in Japan. This would seem to be merely a symptom of a deeper relationship between the Japanese, their robots, and their animated characters -- the distinctions between which are being blurred by this hard-to-pin-down technology called AI.

Hagita, who is also a researcher at NTT's Communication Research Laboratory, said JSAI's goal was the promotion of AI research and the distribution of AI expertise. JSAI was launched in 1986, soon after the creation of the American Association for Artificial Intelligence (AAAI), and currently has about 3,000 members, with roughly 1,000 people subscribing to its mailing list. The response to its Web site has inspired JSAI to look at other ways of providing more accessible information. (See the organization's definition of AI.)

JSAI makes a further distinction between AI research that focuses on "reasoning," such as game playing, and "learning," such as data mining. Other JSAI Web pages include a list of different technologies that might be considered artificial intelligence, such as agent technology or "fuzzy systems" (see glossary of this and other AI-related terms). How do these sub-fields fit into the overall picture of AI? The answer lies at least partly in the history of AI development, starting with the expansion of AI research during the Cold War.

The 5th Generation
The field of artificial intelligence arguably grew out of the Defense Advanced Research Projects Agency (DARPA) funding of American research projects during the '60s. By the early '80s, an expert systems industry had emerged, and both Japan and Europe dramatically increased their funding of AI research. One of the most ambitious projects undertaken was the 5th Generation Computer Systems Project, an attempt to combine European ingenuity with Japanese industrial skill in order to develop a new sort of AI that might rival America's hegemony in the field.

Professor Mitsuro Ishizuka of Tokyo University, the head of the JSAI, was involved in the 5th Generation project and says the idea was for Japan and Europe to create a technology that had no parallel in the US. The project was big-budget (JPY50 billion) and expectations were high -- particularly since it was portrayed in the media as an attempt to produce "intelligent" computers. In the end, it produced the KLIC parallel logic programming language (based on the well-known C language), and some other spin-offs, but nothing of great commercial success. To some, this meant the project was a failure, but as Ishizuka points out, "Research by its nature is not 100 percent successful; if it was, then it wouldn't be research."

5th Generation project technology never really made it into the mainstream, largely because it was beaten to the punch by less elaborate, rough-and-ready techniques such as object-oriented programming. Logic programming, like KLIC, might have had great formalism, and a rigorous theoretical basis, but it was only really popular with AI researchers. The Japanese 5th Gen engineers were applying their expertise to scale this research language up to work with huge, massively parallel mainframe computers, and they got it to work. They thought it would form the basis of the next generation of computing languages, but in the meantime the personal computer had shrunk in size and grown in power, making it far more cost effective for a company to use a cluster of PCs than a huge mainframe computer. By the time the 5th Generation engineers were ready to present logic programming-based supercomputing, the market had out-evolved them. At the same time, the expert system technology that investors had funded in the late '80s was not advancing as rapidly as predicted, leading to a bursting of the AI bubble and the disappearance of many AI startups.

The 6th Generation
The Real World Computing Project (RWCP), aka the 6th Generation Computing Project, is another Europe-Japan collaboration. This 10-year program started in 1992 (around the end of the 5th Generation project) with a much broader mission: The RWCP focuses on a variety of different "softer" technologies that use neural or fuzzy techniques. The research components are much more spread out, and appear to have been selected with an eye for more practical applications of the latest technology. NEC is working on adaptive hardware, Fujitsu on mobile robots, and Mitsubishi on meta-data markup languages, to name a few.

According to Ishizuka, the difference between 5th and 6th Gen projects is that while the 5th Gen project had a clear underlying ideology, i.e. a new generation of programming languages, the RWCP incorporates -- perhaps wisely -- a greater variety of approaches that do not fit under one overarching framework. Perhaps this will be to the advantage of the RWCP. One scientist involved in RWCP work is Dr. Ryohei Orihara, a senior researcher at Toshiba's Knowledge Media Research Laboratory. Orihara's work concerns abstract models for representing sound and images, and their application in interface devices. This framework would allow Web surfers, for example, to easily navigate in a "space of images" in the same way that we currently navigate through text space with Web search engines.

The project is representative of work being carried out under the umbrella of the RWCP, and it shows that the focus has switched to augmenting human abilities, as opposed to replicating them in their entirety. There appears to be a recognition that basic components of human intelligence, such as recognizing images and communicating, are the really hard problems for AI. As Orihara explains, "AI often fails to address the representation problem. This was perhaps its main failing. Recently, representation issues are coming to the fore, and this is to AI's benefit." Still, one of the paradoxes of AI is that expectations keep changing, "making it impossible to implement an AI system," Orihara speculates. "Deep Blue is not AI, ELIZA [a famous chat bot] is not AI, Aibo is not AI; maybe we'll never recognize AI, even when it finally arrives," he adds.

Ishizuka agrees there is something about the term that makes AI an unimplementable technology: As soon as the technology advances, the perspective shifts, and the quality of intelligence passes to those activities that are still only in the human domain. As soon as a computer could beat the world chess master, playing chess was suddenly considered not to require so much intelligence as previously thought, and AI was once again set as an unattainable goal.

Classic AI
So what of the classic AI reasoning systems? Dr. Reijer Grimbergen, an associate professor at Saga University in Kyushu, is researching the application of AI gaming technology to shogi, or Japanese chess. He describes his overarching research objective as "to know more about human intelligence, in particular about the human capabilities for problem solving." Focusing on shogi makes the question a little less abstract. He talks of a special relationship between Japan and AI in terms of the Japanese fascination for robotics and computer games. "Only games that you can play against other human players or against intelligent machines will survive," he predicts.

There is also a downside to the field in general, says Grimbergen. "The problem of working in AI is the controversy that it sparks. Intelligence seems to be something mystical in the minds of most people, and the idea of explaining it is often considered sacrilege." Grimbergen enjoyed the recent Spielberg film and hopes that this will fuel interest in the field. He does not equivocate on the issue of the future of AI: "I strongly believe that this, combined with steady progress in our knowledge about intelligence and the modeling of this, will eventually lead to computer systems that will be considered intelligent without discussion." While Japanese chess might make for more tractable research, one of the criticisms of classic AI was that it focused too much on toy problems, and was not applicable to the real world. Hence the recent shift to fuzzier approaches, and also agent technology, a term almost as confusing as AI itself.

Agent Technology
Agent technology covers a broad range of systems, many of which overlap with AI. The central concept is that of multiple interacting agents, which communicate in order to resolve problems. Gaku Yamamoto, a researcher at IBM Japan, is developing a technology called agent servers, one application of which aims to remove some of the stress from online shopping. The user can specify, for example, something like "I need a 700-MHz PC and printer for less than JPY80,000." An agent handles the query and will contact the user when the right product comes along.

The term "agent" might seem just as vague as AI, and when questioned about these definitions, Yamamoto replies that it's important to think in terms of practicalities. He does not really see his work as AI, since the objective is to make his agent servers as robust as possible. He also disapproves of the hype associated with intelligent agent systems, and suggests that "Real breakthroughs are much more likely to come as a result of agreement on [certain] communication standards." He feels unclear about where agent research might lead, but says, "We'll probably achieve human intelligence in artificial form, but emotions, imagination, passion, vision ... these are not likely to come from an extension of existing technology." The implication is that some radical new approaches are required, and there's certainly no shortage of work on artificial emotions in Japan.

Emotional Agents
Within the field of agent technology is the area of emotional computing. Tomoko Koda, a leading Japanese researcher and entrepreneur, has developed a number of emotional agent systems. She spent time in the US at the MIT Media Lab, working on poker-playing agents that would express different facial emotions depending on the cards they hold. Back in Japan, Koda helped develop the Petaro system -- a very Japanese piece of software: It's an online chat system that allows users to express emotions through character-based faces. The chat interface is wrapped in a cute animal or character of some kind, which expresses the emotional content of the chat message and frees users from confusing conventions like smilies (faces styled out of punctuation marks).

Koda points out that the software is pure emotional expression, with no added intelligence. Petaro currently has over a million users, indicating that smarts are not perhaps as important as getting your target market correct. "AI has no meaning for the average user; it's just something that is not understood," she says.

She also suggests that, to the Japanese at least, "agent" is a more useful term, because it is used in a business sense. The press release announcing the launch of DoCoMo's i-Appli Java service, for example, included the term "agent" (even though it was only vaguely defined). "The user really doesn't care what you label it as, as long as it works," says Koda. "If we get to the point that the system can actually start to deduce things about the user by interacting with them, then perhaps it might deserve the label of AI."

So do we need AI to create emotional agents? Koda's response reflected many of those interviewed: "AI is a concept, not a technology. Real human intelligence is grounded in the stuff that the brain is made of; any emotional system will only be simulating human emotions -- mimicking them. We still can't answer fundamental questions like how many basic emotions a human has, so how can we simulate them?" JSAI's Ishizuka also voiced this opinion, although he pointed out the possibility that real emotional abilities could be evolved using genetic algorithms or learning techniques. We might not understand exactly how the resulting system worked, or whether it really "felt" sad or happy, but either way, a completely realistic emotional interface would have great business potential for customer relationship management.

Affective Computing
Koda's work -- expressive computing -- is all about getting computers to express emotions. But what about the flip side -- trying to get computers to understand human emotions? This research is called affective computing, and Dr. Nadia Berthouze, assistant professor at Fukushima's University of Aizu, is one worker focusing on how humans communicate their subjective states (feelings, to you and me), in an attempt to give computers and robots the ability to understand our emotions. Berthouze explained that emotions are an important part of human decision-making. "There is evidence in neuroscience that demonstrates the importance of emotion in the cognitive process," she says. "Patients who have lost emotive capability [through brain damage or disease] have shown problems in their ability to make decisions."

Berthouze feels that one thing lacking in Japan's plethora of research into affective computing is a strong theoretical basis, and this is one element driving her to develop a set of dimensions that can describe a user's subjective or affective state -- essentially a framework for describing how happy or sad we feel. The field of affective computing and its related emotional disciplines have been labeled by some as the "New AI." As Berthouze's research overview describes, "A car might be chosen for its design possibly because it reflects aspects of the personality of the customer." The emotional framework she envisions is intended to serve as the basis for multimedia data mining, something that would allow computers to help meet, or at least understand, our emotional needs.

Data Mining
A classic example of data mining comes from Super Value, the US department store chain, which found surprising correlations in customer purchasing habits. For example, the firm found that people buying diapers were also buying beer, with the explanation being that those with very young children usually have to stay in during the evening instead of going out to bars, and so were more likely to pick up a six-pack. Super Value increased both beer and diaper sales by moving the two items closer together in their stores. Data mining draws on work from machine learning and statistics, incorporating many techniques that were originally developed as part of AI research.

SilverEgg Technologies (see "Online Marketing With AI," page 55, October 2000) is a Japanese venture company looking to go beyond data mining. CEO Tom Foley explains that their AIgent system observes which product categories a customer clicks on, and then makes intelligent guesses about that customer's preferences. According to Foley, "This is a real, down-to-earth application of AI."

Foley goes so far as to draw a parallel between a business and an "intelligent system," a business having to gather data, draw conclusions, manage data, adapt, and survive. But aren't the intelligent parts of a business the humans that run it? Foley theorizes that a really successful business is best thought of as a kind of framework or intelligent colony that amasses knowledge, and that can outlive the tenure of any particular employee that animates it. Foley raised a striking example. "Consider a business without any employees. It is possible today -- you don't believe it? How about a business consisting only of an adaptive trading program that runs indefinitely, pays an outsourcing company for computer maintenance by credit card over the Internet, and pays shareholder dividends? Is it the computer or is it the business that is intelligent, or both, or neither?" The SilverEgg CEO believes that people are waiting for a machine intelligence similar enough to their own to carry on a conversation, but that if we can see past our anthropomorphized expectations of AI, we may discover that it is already here.

So what distinguishes SilverEgg technology from other data-mining techniques? Isn't it all just number-crunching? Not according to Foley, who points out that number-crunching techniques don't know about the products and customers beyond the correlations they find in the data. In a sense, these techniques spend a lot of time rediscovering obvious relationships -- for example, that books X and Y are correlated when they are in fact written by the same author. SilverEgg's AIgent is designed to enable organizations to use prior knowledge, such as that two books are on the same topic or were written by the same author, to give more accurate and intelligent service.

Interface Agents
So where should we expect this new AI technology to have its main impact? I spoke to Mona, an interface agent running on Inago's NetPeople system, a platform technology for digital human agents. Mona, who looks like Lara Croft in a green metallic costume, can be found on the Inago Web site (www.inago.co.jp). When Mona asked for my name, my initial reaction was to ignore her request, but then I felt guilty and responded. Mona was far more difficult to ignore than an online form -- business implications, anyone?
Ron DiCarlantonio, president, and Ian Wilson, the Japan office AI specialist, explain that Inago's business target is e-business: customer service agents, entertainment, finance, travel, and so forth. The company has offices in both Japan and Canada, following the predictable split of R&D in Canada and marketing in Japan. This situation arose from DiCarlantonio's background as a Canadian who worked on artificial pet software in Japan in the early '90s. Later, DiCarlantonio and Wilson both worked on different kinds of interface agents, but the attitude of potential customers in America was, "My competitor doesn't have it, so why do I need it? I'll wait until somebody else takes the risk." This in spite of the argument that a single animated agent could do the work of 50 telephone operators.
Of course, part of the problem with any new technology is that it needs to be proved in a real business environment. But if no one will take you up on the concept, that may never happen -- Catch 22. Fortunately, the pair were able to draw the attention, and investment, of some major Japanese companies very quickly after showing the NetPeople demo -- even though at the time it was an English-language only product. While it had been very hard in Canada and America, the response in Japan was overwhelming. In America, people said, "This is untried" and "Why do we need a character?", but in Japan, the first thing they said was, "The character makes it," relates DiCarlantonio.
Wilson suggests that understanding this is part of understanding the dot-com implosion: "The driving force behind home shopping was you don't have to go out, but actually people want to have some sort of interaction. If they can see eyes blinking, see a mouth moving, then it doesn't matter that the character is not a real person, but we get the same stimulus, and that's what keeps customers happy." DiCarlantonio suggests that while the current Inago system is not true AI, that's definitely their ultimate objective.
Naturally, Inago's valuation has dropped due to the recent market implosion, but it has established some key business relationships, such as with Horipro, the Japanese talent agency that created the virtual tarento (talent) Kyoko Date. Japanese consumers may find themselves conversing with Date in the not-too-distant future, and when they do, it will be made possible by Inago's NetPeople platform. NetPeople is designed to allow companies to create their own animated agents, with branded personalities like Date, that will handle customer interaction on their Web sites, or over mobile phones. The two AI entrepreneurs look to a future with an open interface specification that will allow programmers to create agent libraries, such as stock characters to handle travel agent interactions or customer satisfaction issues, tell you jokes -- you name it.

Wireless Agents
While Inago is looking to offer mobile characters as one possible interface, other companies are focusing specifically on characters for wireless devices. Tokyo-based I-Chara (see "Social Engineering," page 78, November 2000) is a kind of social networking service for mobile phones where users personalize their profiles and make some or all of that information available to their personal character (think avatar or personal bot). Offline, user profiles are matched with one another (looking for overlapping interests), and they may be introduced to characters belonging to other people or companies.

Dr. Kim Binsted, CTO of eMuse KK, the Irish firm eMuse's new branch that now owns I-Chara, got her PhD at the University of Edinburgh on the subject of computer humor. On the topic of whether intelligence, artificial or otherwise, is an important part of the I-Chara system, she suggests that the crucial factor -- from a business perspective -- is the notion of "apparent agency," the appearance that the character you are interacting with has goals, emotions, and a personality. "You could implement apparent agency in all sorts of ways, but we think the AI approach is the best way to go," says Binsted.

In many ways, creating an intelligent agent has always been the goal of AI research, but the understanding of what intelligence means has shifted. Years ago, solving algebra problems or playing chess were seen as the most mentally demanding "intelligent" pursuits, but as the results of this work led to systems that exhibited little flexibility or common sense, it became clear that the real challenge was replicating social intelligence.

The classic Turing Test exemplifies this. Passing the Turing Test involves convincing somebody that you are intelligent through only a text conversation. Even if a machine succeeds, you could still argue about whether it is merely a "simulation" of intelligence, or if the machine is really experiencing the emotions it might report. But while you do, the people who developed it will be laughing all the way to the bank, since their interface will be handling customers more effectively than anybody else. As Binsted says, "The system has to show social intelligence in the interface."

Cognitive Research Labs (CRL)
So is anybody trying to put all these different technologies together? One of the most sophisticated approaches appears to come from Cognitive Research Labs (CRL), headed up by Dr. Hideto Tomabechi (see "Don't Just Talk to Intelligent Agents -- Be One," page 3, October 2001).

On a conference room video screen at the company's Roppongi headquarters, Tomabechi, wearing traditional Japanese dress and geta (wooden shoes), conducted an agent-powered presentation using MPML (Multi-modal Presentation Markup Language). MPML allows everyday business presentations -- think PowerPoint -- to include instructions about where to position, for example, a Microsoft Agent, what the agent should say, and just how emotional it should get over the subject matter -- alternating from, say, passion or despair.

CRL's work focuses on security issues, with the firm's system integrity based on indestructible data concepts that Tomabechi developed during his PhD studies in computational linguistics. CRL uses a number of artificial intelligence techniques in different products, such as a recommendation system in their online shopping mall, case-based reasoning in their home server system (which allows your video player to learn what a movie is), and decision trees and clustering methods in their banner ad system.

So does Japan have any special relationship with artificial intelligence? Tomabechi suggests that the Japanese are more comfortable with the concept of intelligent non-human entities, because philosophically, Japanese expect all objects to have a spirit; meaning, presumably, that they are less intimidated by the idea of killer robots taking over the planet. So why is CRL "Cognitive Research Labs" and not "Artificial Intelligence Labs"? "The AI boom of the '80s, and subsequent crash, means that AI has negative associations," says Tomabechi, "making the term 'cognitive science' more palatable."

The Emotional Professor
Mitsuro Ishizuka has links with many elements in this story. He was involved with the 5th Generation project, is current head of the JSAI, developer of MPML, researcher in the fields of emotional agents and Web intelligence, and head of the department of Information and Communication Engineering at the University of Tokyo. Despite these high-level pursuits, the professor describes himself as "engineering-focused," with his main goal being to overcome the current barrier of computer interface design. In the '80s, he focused on classical AI systems and logic programming, but by the '90s it had become clear that that approach was not going to be hugely economically successful. Ishizuka broadened his interests into various different types of media, particularly interface agents that could communicate through a variety of modalities -- speech, expression, movement, and so forth. "This may seem trivial on the surface," he explained, "but as soon as you have an interface agent, you are asking it to deliver information, and so up come the concepts of emotion, intelligence, and so forth. This is a mixture of media and AI."

Ishizuka hypothesizes that in the West, a hard boundary is drawn between humans and robots; but in Japan, or even more generally in the East, the distinction is not made so clearly, and robots are seen more as companions and partners. Westerners tend to think of robots as mechanical, whereas in Japan it is assumed they have kokoro, heart and spirit. In Japan, the move to create walking robots has not been hindered by images of terminators taking over the world, and this thinking extends to all sorts of artificially intelligent agents.

Perhaps the lesson here is that if you want to get your information across, if you want to engage your audience or your customers, you need to communicate emotionally. Ishizuka suggests that this is why films are much more popular than technical reports, because they have an emotional message. The real human language is not words, but emotion. A well-implemented emotional markup language would allow people to present their point of view or ideas, along with the emotional information as well. Right now you can do this by giving a lecture and talking about your subject matter with passion, but in the future, perhaps you'll only need to describe your passion, and anybody reading your e-book will get the emotional message.
Whether the agents expressing the emotion will actually be intelligent, or feel the emotions they express, is an open question; but if they can consistently fool us, you can bet we'll see them selling us things and handling our grievances in the near future. And don't be surprised if they occasionally slip into their native Japanese. @