caasnsr@nmtsun.nmt.edu (Clifford Adams) (04/27/89)
[I would have posted this earlier, but our site was off the net for about a month.] About a month ago I finished the following paper for a Philosophy of Technology class. The purpose of the paper was to summarize some works on the philosophy of science and technology, and to relate them to personal experiences. I closely followed the recent comp.ai debate on the Chinese room argument, and tied the agument to my thesis on "real" understanding and learning. My major reference for the ideas in this paper is _The Study Of Man_ by Polanyi, where I first learned of the concepts of personal knowledge. I left the "texification" in the paper, but it should still be readable. I would enjoy any replies, especially replies from the people who originally participated in the USENET discussion. The paper does get better after its initial paragraphs, which are simply an introduction. (at least I hope they do :-) This paper Copyright 1989 by Clifford A. Adams. Permission is granted to duplicate given that this copyright notice is included. ---------- Cut Here ---------- \font\twelverm=cmr12 \twelverm \font\seventeenrm=cmr17 \baselineskip=24pt \parskip=24pt plus 2pt minus 2pt \centerline{\bf \seventeenrm Artificial Intelligence and Learning} \centerline{\bf \seventeenrm A Changing Perspective} \medskip \centerline{\rm by Clifford Adams} \bigskip I have had an interest in Artificial Intelligence for many years. My interest began when I played with Eliza, the program which simulates a Rogerian analyst. It seemed to have a certain kind of ``life'' to it that other programs didn't have. I was puzzled because I knew exactly what instructions were in the program, yet the correctness of most of Eliza's responses was amazing. I soon started adding features and information to Eliza in hopes of having real conversations with it. My additions added very little to the sense of really conversing, and I soon abandoned the project. At the time, I saw Artificial Intelligence as a field that simply needed to have problems solved. When the problems were solved, intelligence would arise. Soon after I arrived at Tech I learned about many spectacular programs which solved many of the interesting problems. I thought that the AI community was most of the way toward solving the needed problems, and that intelligent computers would soon become a reality. Then I noticed that no real or measurable progress had been made toward the final goal of AI. I learned about the Turing test, and believed that any computer able to pass that test must surely be intelligent. ``No problem,'' I thought. ``all that's needed is a bit of integration of the solutions.'' The integration then became a hard problem. It was at this point that I decided not to make AI the core of my Computer Science major, although I still have a strong interest in the field. The first part of this paper will explain how my views on AI have been affected by the works of Heiseberg, Kuhn, and most importantly Polanyi. The remainder of the paper explains how Polanyi's writings on understanding has changed my views of learning and comprehension. Artificial Intelligence has been a field deeply divided by competing paradigms. There are no unifying theories or experimental procedures in the field. One of the greatest problems is that researchers often do not agree on what AI even means! This is unlike other fields, like chemistry and physics, where many agree that a problem belongs in the field. Anything from problem-solving to creativity to Barry Kort's ``ethical systems'' is a research topic for someone in the field. The most obvious split in the field is between those who attempt to recreate natural intelligence and those who attempt new approaches. Natural intelligence arises from neurons, so those in the ``natural'' camp often attempt to discover how these neurons are organized to create intelligence. They then try to organize ``neural nets'' of simulated neurons in order to recreate the processes. The results have been amazing for small problems such as recognition or ``learning'' of a specific type of action, but no large-scale reasoning processes have been simulated successfully. The second camp, which I have more interest in, is the creation of intelligence by means that are more ``artificial'' than neural nets. This is also called the symbolic approach to Artificial Intelligence. The idea behind this method is that intelligence can be modeled or created by formal means. This kind of view is often held by physical scientists such as Heisenberg, who believe that the universe can be explained by Platonic formal symbols. Polanyi argues in {\it The Study of Man} that the processes of understanding and intelligence are not monolithic. He shows that two kinds of understanding are important: both explicit, rule-based understanding and tacit, experience-based understanding. Polanyi then states several times that the tacit understanding is a personal experience which cannot be done with strictly formal operations. This conflicts with the view of many AI researchers who believe in ``strong AI''. They believe that it will some day be possible for a machine to have ``true'' understanding in the same sense that people do. Polanyi says that such a goal is not reachable, because formal operations are inadequate for tacit understanding. One reason for the impossibility is the total lack of a difference between Polanyi's ``focal'' and ``subsidiary'' awareness. The machine cannot focus on the whole without focusing on the parts. A computer either knows something or it does not--there is no ability to focus on a comprehensive whole while having a subsidiary awareness of the particulars. More basically, because computers lack the ability to experience (as opposed to simply recording), they lack the foundation of tacit understanding. Therefore the computer cannot understand as people understand. Various thought experiments have been proposed as tests for the existence of intelligence or understanding. One of the tests, the Turing test (named after Alan Turing) is often cited by AI researchers as being a good test for understanding. The idea is that if a person cannot tell the difference between conversations with a human and conversations with a machine, the machine has demonstrated intelligence equal to a human being. The test appears to be correct since it seems to require human intelligence to converse with another person. The main argument against the Turing test is the Chinese Room problem posed by Searle. He said that a person could be placed in a room, given explicit rules of action which do not require understanding (such as ``write squiggle when you see loop''), and given Chinese text to process. If the rules are correct, the person in the room will write good Chinese answers, but will not understand the questions or Chinese in general. The Chinese Room argument is one of the most misunderstood [sic] arguments in Artificial Intelligence. The room behaves as if it understands, yet it does not understand. Some proponents of the Turing test say that if the machine acts as if it understood, then the machine actually did understand. This is due to the confusion between explicit and tacit understanding. It may be fair to say that the Chinese room {\it explicitly} understands, but it definitely does not understand tacitly. The Chinese room argument states that it is impossible to know by symbolic means that a machine is doing anything (tacit understanding) but manipulating symbols. One enhancement to the Turing test which defuses the Chinese room is called the Total Turing Test, by Stevan Harnad. Harnad proposed a test in which a computer is placed inside a robot which looks and acts exactly like a human being. If a person cannot tell the difference between the robot and a human, even after extensive interaction, then one must conclude that the robot is truly intelligent. The reasoning behind this is simply the Other Minds problem--if you can't tell the difference between the robot and a human, and the human is assumed to have a mind, then the robot should be considered to have a mind. Learning is one of the most important facets of understanding. Learning is the process by which a (person {\it plus} knowledge) becomes a (person {\it with} knowledge). The ``plus'' indicates that the person has the knowledge in some form (such as the rulebook in the Chinese Room), but does not tacitly understand it. When learning occurs, the knowledge becomes an integral part of the learner. Explicit knowledge is easy to forget, but tacit knowledge changes the owner of the knowledge. Tacit knowledge is often hard to develop. Learning to ride a bicycle is a common process of tacit learning. None of the advice or suggestions really make sense until one ``gets the idea'' (learns) how to balance. Another example I am experiencing is in folk dancing. The moves often happen too quickly for much thought to occur, so the patterns must be learned tacitly, by ``teaching the feet, not the head''. Polanyi also recognizes that conscious thought may not be appropriate for some learning situations. Public education has exposed me to many new experiences and requirements. Most of what I remember has been that which I have learned both explicitly and tacitly. I obtained little value from some courses that I didn't like. For those courses I only ``learned the tests'', which was exclusively a process of explicit understanding. Some courses (especially math) I had a ``feel'' for, but was unable to solve problems because of a lack of explicit understanding. I experienced both problems in a Spanish class. I learned the language explicitly for the first year, but I didn't really understand the language. My thoughts were closer to the man in the Chinese room than those of a native speaker. Then I suddenly understood the language. I could think in the language and speak Spanish without mental effort. Unfortunately, I knew very little of the language at that point, and the novelty of thinking in two languages soon decayed, to the point where I can only remember what it was like to think in another language. I have also noticed the differences in understanding at Tech. Many people who choose Computer Science as a major don't know that a large amount of tacit understanding is required. Programming a computer is like speaking in a language. The transfer of ideas between languages must occur smoothly and quickly in order to program efficiently. Some people seem to believe that they can simply program ``by the rules''. They are then surprised when they need to design and write a program by themselves, a nearly impossible task without a tacit understanding of computers and programming languages. Those who understand can simply write code without specific thought, with a focal awareness of the code segment, and only a subsidiary awareness of particular lines or statements. Pieces of code begin to look ``good'' or ``bad'' to a programmer depending on whether they fit her mental models. Errors in the program quickly become obvious, instead of needing an exhaustive search. In general, the tacit understanding is what makes computer programming an enjoyable activity, rather than a tiresome chore. In conclusion, Polanyi's work has made a major change in the way I think about understanding. I have had a tacit understanding of many of his views, but his papers allowed me to explicitly discover the difference between tacit and explicit understanding. I also now realize that this paper is mostly a test of whether I have tacitly understood the texts used in class or if I am just parroting explicit knowledge. \bye -- Clifford A. Adams --- "I understand only inasmuch as I become." caasnsr@nmt.edu ...cmcl2!lanl!unm-la!unmvax!nmtsun!caasnsr (505) 835-6104 | US Mail: Box 2439 Campus Station / Socorro, NM 87801
bwk@mbunix.mitre.org (Barry W. Kort) (05/01/89)
I enjoyed reading Clifford Adams' essay, "Artificial Intelligence and Learning--A Changing Perspective." I find myself increasingly amused by the debate over "True Understanding." After reading Feynman's anecdotes about Brazilian physics students (in _Surely You're Joking, Mr. Feynman_), I now differentiate between shallow understanding and deep understanding. I believe there is no theoretical limit to the depth to which one can achieve understanding. I find the word "understanding" a bit problematical, because I don't fully understand what the word means. I prefer the word "comprehension" because its etymology is clearer. "Comprehend" means "to capture with". I capture knowledge by constructing a mental model that resembles (in both structure and behavior) the object of my contemplation. Often, I need to construct a physical model to play with, or a computer model to interact with before I can get a good mental picture. Using Lisp as a metaphor, I measure the depth of my understanding by the number of levels of "chunking" or decomposition between my overall model, and its atomic constituents. I also measure my depth of understanding by the number and richness of concrete instances of an abstract model. Here is where computers and AI are a bit behind humans. When it comes to models, analogies, metaphors, and parables, computers are just getting started. Symbolic representation is in its infancy. For me to understand something well, I need multiple interchangeable models. I need a verbal or mathematical representation, which I can manipulate formally, and I also need a visual or geometric representation which I can manipulate like a cartoon in my head (or on my computer graphics display). To my mind, the next great advance in computer understanding comes with the computer's ability to project an animated color image that graphically represents (re-presents) the structure and behavior of a system by transforming a symbolic (i.e. ASCII) representation into a visual (or audible) form. Bi-directional information-preserving transformations are the central tool of modeling. We see examples in Fourier Transforms, Analytical Geometry, Digital Signal Processing, Holograms, Analog Computers, and various Duality Theories. --Barry Kort