SCIENCE

MENLO PARK, CA -- Anticipating this summer's International Joint Conference on Artificial Intelligence (IJCAI-01), to be held August 4-10, 2001 in Seattle, Washington, four of today's leading AI experts gathered to discuss some important issues. The participants: Howard Shrobe is a Principle Research Scientist at the MIT Artificial Intelligence Laboratory (MIT AI Lab). He served as Associate Director of the Laboratory for the past three years. He served as Assistant Director and Chief Scientist of the Information Technology Office at Defense Advanced Research Projects Agency (DARPA), and was responsible for the Intelligent Systems and Software Technology group. He also helped to design the White House's electronic publication system. Bruce Buchanan is the current AAAI president and Professor of Computer Science, Philosophy, and Medicine with the Department of Computer Science at the University of Pittsburgh. His research interests are machine learning, knowledge-based systems, medical expert systems, and computational biology. He has studied applications of machine learning and artificial intelligence as they pertain to biology and medicine. Tom Mitchell, the incoming president of AAAI, is currently on a leave of absence from Carnegie Mellon University to work at WhizBang!, a company that applies artificial intelligence and machine learning to automatically extract information from all over the Web. His extensive publication record includes looks at machine learning and robotic intelligence. Tucker Balch is a research scientist in robotics at Carnegie Mellon University. He is interested in all aspects of building and operating large-scale multi-robot teams that require cooperation, diversity, communication, distributed learning, and distributed planning. A former fighter pilot, Balch has worked as a computer scientist at the Lawrence Livermore National Laboratory and in the Robotic Vehicles Group at the Jet Propulsion Laboratories. Q: What misconceptions exist about AI? Mitchell: It seems people generally associate AI with the grand proportions of human-like machines with human-like characteristics popularized by fiction and the entertainment industry. People bring those preconceived notions to their understanding of AI as a technical and scientific discipline, and many assume that AI scientists are working to build robots that look and act exactly like humans. But that is closer to fantasy than reality. What the field of AI is really about is inventing machines that will help people in a variety of ways, by giving machines some of the sophisticated capabilities that humans have, such as the ability to understand spoken words, or interpret images, or to learn from experience. Usually these machines do not look or act at all like people, but they can be amazingly useful to people by improving and assisting our lives, and complementing rather than replacing the things that we humans like to do. And that's the goal we are collectively working toward. Q: How do we make this real? Mitchell: It is real now. We already have computer systems that understand speech, and are being routinely used by the phone company to provide telephone numbers when you dial for information. We already have computerized robot systems that handle many dangerous manufacturing tasks, for example welding tasks in automobile manufacturing. We already have computer systems that learn from historical medical data to predict which medical treatments will work best for which future patients. There are many examples of AI systems that have one or two human-like capabilities, which are in routine use and contributing to our economy and our quality of life. Balch: A byproduct of popular fiction is that the public may have an unrealistic perception of what machines can do today. When they interact with a real AI system they are often disappointed. However, we are making tremendous progress in the field in step with the advances in technology happening around us. The reality is that AI is good at some specific problems, like chess, data mining, and knowledge management, but we don't yet have the "whole iguana" so to speak. The nature of research is to work devotedly to solving and improving one technical challenge at a time and secure successful integration. We are working to build reliable systems, one critical piece at a time. Q: What are the challenges that keep it from us having the "whole iguana?" Balch: Much of it has to do with hardware. We still need faster computers and larger data storage capacities before we'll really start reaching animal-level intelligence. Shrobe: The combined systems of computers and people will be more capable than computers alone or people alone. And that's really the Holy Grail for us: to use our achievements to enhance the daily lives and tasks for humans. Q: Must AI products take human form? Balch: No, not at all. When people think of "AI," they generally seem to think of machines in human form. But, we have robotic dogs and robotic soccer players that are shaped like little boxes. AI can be "behind the scenes," too, for example, the paper clip in Microsoft Word. Mitchell: Of course they don't have to be human, but even if they don't look like people it can be very helpful if they communicate like people. All of us know the frustration of trying to communicate with computers by typing in arcane commands or trying to find the right system parameter to reset. It would be so much better if our computers could simply accept our commands in English or other natural languages. That is one reason why our field puts a lot of effort into human-like modes of communication such as natural language processing, vision, etc. Some of the latest research on robotics has also gotten into interesting issues of how to communicate with robots via gestures. Buchanan: Human form is incidental to our understanding intelligence. It controls the way we interact with the world, even though there are sometimes better ways of doing things. For example, moving with six legs is more stable than moving on two. However, having robots walk on two legs can help us understand our own limitations. What we find is that we are learning as much about our own human systems and the limitations of biology as we are about computer systems. Shrobe: Being close but not the same could be worse than being far away. A human-like system doesn't need to be embodied as a human, like Hal in "2001." And while the human form may be useful, it may be better to build something that can assist the elderly rather than something that is more "human-like." Q: Science fiction offers us a depiction of advanced machines that begin to take on characteristics of emotions gone out of control. Should we worry about this? Mitchell: No. This won't happen in any near time frame. But you can make these kinds of statements about any technology, including biomedicine, computers, even automobiles. You just pursue the technology in a reasonable way. Balch: We shouldn't be afraid of AI because humans are much smarter than any program. There are, of course, different aspects of being "smart," but we won't need to worry about machines taking on the embodiment of autonomous emotions. This will be true for a long time. The misconceptions hurt because they build up unrealistic expectations about what AI and computer science can do. Q: What's next? Buchanan: In the next two years we will be seeing mobile robots that are capable of doing more simple, limited tasks. They will provide assistance in the kind of rapid decision-making that may be difficult for a person, like maneuvering through traffic, and intelligent assistants that you can use around the house, like a firefighting robot with a built-in vacuum cleaner that can suck up flames and smoke. We'll continue to see more and more programs that provide intelligent help with the computers, cars and household machines in our lives. Shrobe: I'm very interested in understanding why organizations repeat their mistakes. I believe that an AI system could help capture design choices by looking at what people sketch and listening to what they say. By understanding these common modes of interpersonal communication, a computer could have a much better idea of what people were thinking about when they made these choices. Right now it is very hard to capture this information, and it is not immediately relevant to what designers are doing. So it's more trouble than it's worth. But if we can use perceptual technologies to help capture these design choices, we will make the capture of design rationale and other decisions cost less and have them happen automatically. Q: What do you expect to accomplish at this year's conference in Seattle, and others like it? Shrobe: Science doesn't make progress by great geniuses sitting in rooms by themselves. Progress happens when people bounce ideas off of each other. Great discoveries don't come from an individual; they usually come when a scientist is musing over someone else's ideas. The work of organizations such as the AAAI and IJCAI create forums for dialogue, highlight and recognize achievements in the field, and help in general terms to spur the progress of the science of AI. Buchanan: Every conference has multiple goals-to engage young people, highlight some of the best work, illuminate the problems, and point to partial solutions. It also affords the public a chance to drop in and see AI solutions in action, hear about the fascinating work of their researchers and authors of great ideas, and consequently be inspired themselves toward a better understanding of the field. IJCAI-01 is sponsored by the International Joint Conferences on Artificial Intelligence, Inc., the American Association for Artificial Intelligence, AT&T Labs, Boeing, SemanticEdge Technologies, Microsoft, and NEC Research. IJCAI-03 is slated for Acapulco, Mexico. For registration go to: http://www.ijcai-01.org Founded in 1979, the American Association for Artificial Intelligence (http://www.aaai.org) is a nonprofit scientific membership society devoted to advancing the science and practice of AI. Its mission is to: (1) advance the scientific understanding of the mechanisms underlying intelligent thought and behavior, (2) facilitate their embodiment in machines, (3) serve as an information resource for research planners and the general public concerning trends in AI, and for the general public, and (4) improve the training of the current and the coming generation of AI researchers and practitioners.