|
|
|||
The Secret Agent Man
In a 1997 collection of essays entitled HAL's Legacy: 2001's Computer as Dream and Reality, leading computer scientists and philosophers commemorated HAL's 30th birthday as a movie-star by exploring this very question. These authors debated how far we are from achieving a system like HAL - or even if such an intelligent agent is possible. Roger Schank, the director of the Institute of Learning Sciences at Northwestern University, argued that "HAL could never exist" because "We cannot expect computers to attain understanding that is beyond whatever level knowledge they possess." "That is the bad news," he wrote. "The good news is that many AI researchers have become sophisticated enough to stop imagining HAL-like machines," remarked Schank. Nonetheless, some AI researchers, like Doug Lenat, argued in the book that intelligent agents can be developed that embody the level of understanding demonstrated by HAL's natural language conversations in 2001: A Space Odyssey. Lenat is the chief architect of a project dedicated to building just such an agent, called CYC. Lenat's company, Cycorp, is attempting to give CYC a conceptual understanding of the world by using second order predicate calculus to create hundreds of knowledge bases or microtheories. Lenat hopes that these microtheories or contexts will eventually enable CYC to reason like a human (i.e., using common sense), communicate in natural language, and ultimately learn on its own. Originally, in 1983, Lenat thought it would take 10 years for CYC to achieve these loft goals. However, over the years Lenat has had to revise his timeline on numerous occasions. In the 1997 Wired Magazine article, Happy Birthday, Hal, Lenat speculated that by 2001 CYC will be able to learn English on its own, unsupervised. However, in a more recent conversation, Lenat told me that CYC was still unable to communicate in natural language (i.e., conversations much be carried on in second order predicate calculus - talk about a party stopper!). Regardless of whether or not Lenat meets his latest self-imposed deadline, it is clear that understanding is the key to Natural Language Processing (NLP) and the way to achieve understanding is through knowledge. What is not so clear is how to endow a computer with knowledge. In the years since the formal birth of AI (in 1956) the field has splintered into separate camps. A significant group of scientists now explore and advocate approaches that more closely parallel the natural structure and function of the human brain, while the traditional AI researchers persist in their efforts to use symbolic methodologies to represent and process knowledge. Current natural-based approaches include neural networks, genetic algorithms, and other brain-like software adaptations. Bayesian belief networks are the dominant symbolic methodology, according to Stuart Russell and Peter Norvig in their seminal textbook Artificial Intelligence: A Modern Approach (Prentice Hall). Russell and Norvig claim that, "The belief network formalism...largely overcomes the problems with probabilistic reasoning systems of the 1960s and 1970s...." by specifying dependence between variables using joint probability distributions. Belief networks actually function much like neural networks, but to relief of the symbolists, provide traceable results, unlike multi-layered neural networks. The significance of belief networks is underscored by the interest of Microsoft Research (MSR) in this knowledge representation technique. The Decision Theory & Adaptive Systems Group within MSR is actively developing and using belief networks. These researchers have even made available a downloadable Windows application, called MSBN32.EXE, for developing your own belief networks, as shown in Figure 1. This application lets you load and store belief networks in textual form, create and modify networks through the addition of nodes and arcs, assess discrete probabilities, and evaluate belief networks using exact clique-tree propagation methods. (MSBN32 was written by David Hovel, at MSR, and can be downloaded at no charge for non-commercial research and educational use at http://www.research.microsoft.com/msbn/default.htm.) Microsoft is betting big on the emergence of an HAL-like computer. Bill Gates, in a speech at the CA-World 1998 technology conference in New Orleans, said "I'd be so bold as to say that 10 years from now every personal computer will have seeing, listening, and learning." Gates believes these advances are reasonable because researchers now understand that "It is only because of common sense and context that people are able figure out what's being said." Gate's remarks are not merely idle conjectures, the Natural Language Processing group at Microsoft Research is aggressively recruiting people with AI and linguistic skills in an all out attempt to make his visions reality. However, creating natural language processing systems is easier said then done. It is one thing to build speech recognition and synthesis systems (which are now 98% accurate), but quite another to develop an agent capable of true NLP. Hubert L. Dreyfus, Professor of Philosophy at the University of California, Berkeley, is an outspoken critic of the possibility of endowing a digital computer with language abilities. Dreyfus argues convincingly in his book What Computers Still Can't Do (MIT Press) and in a subsequent paper Response to My Critics that, "To learn natural language a computer has to have a body; it must be embodied if it is to be embedded...to learn a language is not just to learn a fixed set of words and grammatical constructions, but to use this linguistic equipment in ever new situations...it is this ability to project a language into new situations that shows we have understood it...we can't do this projecting without appeal to bodily analogies that sense directly because we have to move, overcome opposing forces, get a grip on things (and ourselves), etc." Eric Horvitz, the program manager (and senior researcher) for the Decision Theory and Adaptive Systems Group at MSR, firmly believes that the AI community is now on the right track in developing machines that will be able to "understand" language. Horvitz quipped that AI critics like Dreyfus are much like the folks that decried the possibility of human flight around the turn of the 19th century. Others, like Simon Corston-Oliver and Bill Dolan (researchers in the Natural Language Group at MSR), argue that computers really don't need to truly understand language to engage in meaningful and useful conversations. Corston-Oliver and Dolan said that sophisticated linguistic knowledge representation and processing systems under development at MSR, like MindNet, will eventually enable computers to interpret language well enough to create the illusion of comprehension (see Figure 2 for conceptual view of how MindNet captures the meaning of language). So, will we one day have the ultimate HAL-like chat bot capable of engaging in riveting conversations about the world and working tirelessly to perform a multitude of actions on our behalf? In a recent interview with Ramanathan Guha (the former co-director of the CYC project), he told me that his seven years spent on CYC was a lot like trying to build a starship with 17th Century technology. However, as our tools mature, even Hubert Dreyfus expects, "...that there will someday be androids like Data in StarTrek: The Next Generation..." One thing is for sure, if an HAL-like entity is built, we should make darn sure that turning him (or her) off is an easier task than what Dave Bowman had to undergo in the movie 2001. This is 007.08 over and out. Until the next transmission, you can reach me at don@donbarker.com or visit me at my Web site (http://www.donbarker.com). Just remember that if you let the wrong word slip while kissing persuasive lips, the odds are you won't live to see tomorrow.
|
|||
It's Almost 2001 - Where's HAL?
