The Machine That Changed The World
PBS Video Series
A production of WGBH and the BBC
Topic Summaries & Notes
© 1997-8 by David J. Stucki
Many of the links provided below provide more detailed or in
depth information about the highlighted topic or person. Others may lead
you to only loosely related sites that are worth exploring.
(Thanks to John
A. N. (JAN) Lee for providing links
to many of the resources listed below.)
The
Thinking Machine, Episode 4
Introduction
How are people different from computers? Can computers think? The answers to these questions, although seemingly obvious to children, have taxed the minds of some of the greatest thinkers of the 20th century.
Although computers were originally designed to serve as calculating machines, many of the earliest computer scientists and engineers thought of computers as thinking machines. As far back as the 1950's there was work on developing computer programs that could play checkers, or
chess, or solve calculus problems.
Thinking, however, is a mysterious activity. Scholars have often described thinking not as a process of the brain, but one of the mind. Pioneers of artificial intelligence (AI), following this philosophy, had little interest in how the brain was constructed. They saw the problem of artificial intelligence as that of simulating the mind, an endeavor entirely different from simulating the brain.
A strong analogy was made, identifying the brain as the hardware (or wetware) upon which the software, the symbolic processor of the mind, was executed. This metaphor allowed researchers to take the position that the solution to the problem didn't necessarily have to mimic nature's solution, just as in the example of manned flight not involving airplanes that flap their wings to remain aloft. We could bypass the biology and evolution and directly engineer intelligence.
The Golden Years
In 1958 Marvin Minsky and John McCarthy formed an AI department at MIT.One of their first students, Jim Slagle, developed a program that could solve college level calculus problems. The program consisted of less than several hundred rules governing symbolic integrals, yet it was able to get an 'A' on the MIT calculus exam.
Hubert Dreyfus, a philosopher and AI critic, recalls students coming to him talking about philosophy having spent centuries on the issue of mind and making little progress, while in only a very short span of time they had practically finished understanding all of the aspects of cognition through the model of computation. He comments that although initially he had no idea what they were talking about, he soon became interested in investigating their claims for himself.
In general, the decade of the 50's was one of optimism. The technological breakthrough of the computer was still so new that there were no visible limits to what could be accomplished with computation. Oliver Selfridge, a respected scientist, in an interview said that he thought that machines would be able to think, at least that they would be able to do things that we would describe as thinking when done by people. This optimism persisted, for the most part, until a scathing report was published by Sir James Lighthill in 1973 criticizing much of AI research.
One of the early successes in AI was in the area of microworlds (also called toy problems). For example, computer systems that were designed to manipulate a collection of blocks on a table by means of a robotic arm. Or Edinburgh University's Freddy system that was able to recognize objects placed on a round board. These early systems faced serious problems that would continue to haunt AI for years. The block's world system lacked the common sense to build a stack from the bottom up. It had no knowledge of gravity. Freddy required over ten minutes to recognize a cup: a task a child could accomplish instantly.
Many researchers began to gravitate away from perceptual problems into a more abstract domain. The computational issues in areas such as vision were too immense to work with. An example of this is the Stanford Cart project from the early 70s, which attempted to navigate across a field of obstacles. It required a fifteen minute pause between each meter traversed.
The Disembodied Mind
In 1950 Alan Turing wrote a paper describing what is now known as the Turing Test. The idea was that if a computer could be programmed to engage in conversation in such a way that it could fool someone into thinking it was human, then it must be intelligent. Early systems inspired by the Turing Test were much too brittle to be considered intelligent. Take, for example, Joseph Weizenbaum's program Eliza. Eliza used simple text processing and keyword matching techniques to simulate the kind of discourse common in psychotherapy. Because Eliza did not analyze more that the superficial surface structure of the sentences, it was easily unmasked as nothing more than a parlour trick.
Without experiential common sense or general background knowledge of the world, a computer will always struggle to discern the appropriate context that it should bring to bear on understanding a given situation. Doug Lenat provides an example of ambiguity that must be resolved by general background knowledge in a story in which a girl sees a bicycle through a store window. We are told that she wants it. The question becomes how do we know that it refers to the bicycle and not the window or the store?
Without the ability to apply contextual or general knowledge to a situation a computer will never be able to fill in the detail required to completely, or even adequately, understand a body of text. This became painfully obvious in the area of natural language translation. Although early predictions indicated that technology was on the verge of providing Russian-English translation systems that would accomodate the entire technical written output of the Soviet-Union in a few hours of processing a week, the actual results were much more disappointing. (One story told that a system that was developed for translation in both directions was tested by translating English->Russian->English and comparing the result with the original. The quote, "The spirit is willing, but the flesh is weak," came back as, "The wine is good, but the meat is spoiled.")
One of the roles that common knowledge plays in helping us to understand natural language is to resolve its ubiquitous ambiguity.
By the end of the 60s, it was becoming clear that the problems AI had tackled were much more complex than had at first been thought. Marvin Minsky observed that the things we have traditionally associated with intelligence, things we find to be difficult, are relatively easy to program computers to do, whereas those things that are so easy for us that we often take them for granted are exactly the tasks that are the most difficult to program computers to do. With the publication of Perceptrons by Minsky and Papert in 1969, What Computers Can't Do by Hubert Dreyfus in 1972, and the Lighthill report in 1973, the future of AI looked bleak and much of the funding of research was curtailed or killed.
Expert Systems
Researchers did not give up, however. One of the significant advances during the 70s was the development by Terry Winograd of a system called SHRDLU. SHRDLU was a microworld system that was successful at using natural language to converse and follow instruction with respect to a block's world environment. Because it could only operate within a limited domain, people outside of AI were not impressed, mainly due to the brittleness of systems that operate in a narrow domain. Edward Feigenbaum, on the other hand, saw an opportunity to produce systems commercially that could serve as experts in selected domains: expert systems. He developed an expert system, DENDRAL, that was able to perform structural analysis of organic compounds based on mass spectrometry data. The commercial applications of expert systems helped to shore up AI research through the dark years of the 70s.
Expert systems are comprised of the codified knowledge of human experts, who were shocked and dismayed to learn that they had devoted their professional lives to tasks that turned out to be nothing more than the application of a few hundred rules.
Although expert systems provided a renewed opportunity for research, many critics also found a new opportunity for condemning AI. Hubert Dreyfus points out that although an expert system for blood disease diagnosis may be able to outperform some human professionals, it doesn't even understand what a germ is. Doug Lenat (not a critic, but one of the foremost proponents of AI) describes a system that grants auto loans to 19 year olds who claim to have had 20 years experience at their jobs and a system that diagnoses an old car to have measles. The message is that systems that have deep but narrow knowledge are extremely brittle, and can't really be considered to be intelligent. Similarities can be drawn between such systems and human idiot savantism. For example, some individuals can tell the day of the week for an arbitrary date, past or future, but are unable to do simple arithmetic.
Commonsense Knowledge
Work also began in the 70s to address the brittleness AI systems. One approach was to build a collection of frames or scripts that captured the stereotypical features of a scenario or event. Then once a story was identified with the appropriate script or frame, the system would be able to fill in any details missing from the narrative. This approach worked fine until there were details missing that did not appear to be appropriate to any script or frame. Thus was identified the commonsense (or general background) knowledge problem.
How do we accomodate discourse that assumes knowledge that is common to all people of all cultures, but that is so obvious to us that we never would have thought to tell the computer? Things like: children are never born before their parents, if you are running a marathon you are awake, or if you fly to Denver the other passengers on your flight also go to Denver. Marvin Minsky has estimated that there are approximately 10,000,000 items or units of general background knowledge, although Dreyfus claims that the number of items is practically unlimited.
General background knowledge became the nemesis of AI as we entered the 80s, a severe obstacle to progress. In response, the focus of interest shifted from knowledge engineering to machine learning.
CYC
Doug Lenat began CYC in 1984, a ten year project, with the goal of codifying all of the general background knowledge necessary for computer systems to intelligently engage in natural language communication. The idea behind CYC was that it would catalog information like an encyclopedia, but exactly that information that you would never find in any normal encyclopedia because it is so obvious and widely known.
One of the principles upon which his project was based is that learning only occurs at the fringe of what you already know, and that the more you know the more and more quickly you can learn. His contention was that AI systems were generally too impoverished for any kind of sophisticated performance.
There were, however, other approaches to machine learning that did not seem to be concerned as much with background knowledge. One of these approaches was based on the idea that common sense relies primarily on something no computer possesses: a human body. In other words, common sense is not based on facts, but on experience and skills.
Connectionism
From the early 50s there have been ongoing attempts to build artificial brains. The computational models used are called neural networks, and the researchers who work with them are connectionists. A neural network learns by adjusting the strengths of connections between neurons, weighted patterns of neural discharges. An early neural network, the perceptron, was popular in the 60s. One application that was attempted with perceptrons was distinguishing male from female faces. Interest in neural network research dropped significantly in the 70s, after Minsky and Papert published their analysis of the weaknesses of the perceptron.
Connectionism experienced a resurgence in the 80s with the development of more powerful neural networks based on the backpropagation learning algorithm, which overcame many of the weaknesses of the perceptron. One of the most successful projects was Dean Pomerleau's NAVLAB project at Carnegie Mellon University. In the summer of '94 a neural network navigated a vehicle "No Hands Across America" from Pittsburgh to San Diego.
Neural nets, however, have problems of their own. There is no way to "look inside" a neural net to see what it has learned. Sometimes what they actually learn is not what you think they have (or what you would want them to). Hubert Dreyfus tells the story of a net that mistakenly discriminated photos based on which were taken on sunny or cloudy days, rather than the intended difference of the presence or absense of army tanks in the pictures.
Another successful (and perhaps the most famous) neural net, Sejnowski's NetTALK, uses backpropagation to learn to convert text input into pronounced English audio signals. Initially the net makes a random attempt at pronunciation, but its output is compared to the correct output, and the difference (error) is fed back through the net (backpropagated) to effect a change in the net's connection pattern. NetTALK understands nothing of the text it pronounces! Researchers have determined that the brain is not a single, general purpose machine, but rather an integration of many specific modules: vision, language, etc.
Future Prospects
Artificial Intelligence has developed many technologies that have been developed into commercial applications. Among these are chess playing programs, expert systems, airline scheduling systems, hospital robots, readers for the blind, and software for limited translation of Japanese into English. All of these, however, would fail the Turing Test. They don't achieve the quest for a general purpose intelligence.
So is there any hope for the future? Some have suggested that the last best hope is the continuing work of the CYC project team. Doug Lenat claims that his prospects for the success of CYC are greater than two thirds (six times what he would have predicted in 1984). Hubert Dreyfus speculates that if CYC does not fulfil its promises, and never is able to understand the kind of stories understood by four year olds, that AI will be dead for however many hundreds of years it takes for us to figure out what intelligence really is.
Comments, feedback, or inquiries, please contact DStucki@otterbein.edu
