AI: The Technological Trickster
By ai-depot | October 26, 2002
Do Computers Think or Do We Just Think They Don’t?
The foundation of information theory as it applies to computer systems derives from the work of Claude Shannon at Bell Labs in the 1940s and 1950s. Shannon’s theorems suggest: (1) a message must contain enough redundancy to survive transmission through noise intact and (2) it may utilize a code for that purpose. (Campbell, p. 113) Using this basis and analysis of message transmission, he was able to develop testable hypotheses of the amounts of entropy within information systems, leading to equations and theorems. These are a powerful conceptual tool for exploring the parameters of both human and machine intelligence in the ability to process information.
With Shannon’s concepts, Waldrop suggests “… an AI researcher can in principle …put human intelligence in a wider context in much the same way the Wright brothers put sparrows and eagles into a wider context.” (Waldrop, p. 30) Information processing may be something that takes place in any system above a certain level of complexity that interacts with a larger environment. Intelligence may be that interaction operating on itself in some way that produces, at most, self-awareness, and at least, an ability to learn from experience. Both require at least an ability to manipulate some understanding of a world beyond the self, if not parts of that world itself, or symbolization.
Building on Shannon’s work, Herbert Simon discovered that computers are able to handle symbols in the widest possible sense. Simon originally wrote on goal-directed behaviours in corporate organization, suggesting humans “…tend to identify with sub-goals rather than ultimate goals…” relative to our areas of expertise, and that we don’t maximize our goals, rather we “satisfice.” We accept the best compromise among identified alternatives. Simon discovered at RAND “…the computer was being used not to generate numbers, but locations - points - on a two-dimensional map.” He recognized that the computer is a general symbol manipulator, ultimately “…a medium for working with ideas…” (Waldrop, p. 21�22, 203)
Morton Wagman has studied the EURIKA program developed by Simon and his co-worker, Allen Neville, to analyze the logic theorems of A. N. Whitehead and Bertrand Russell, and even design some new proofs. He concludes the program adds information to its storehouse (not memory) by plugging a particular set of inputs into a formula, and that while this may be reasoning, it can not be thought of as creative reasoning. (Wagman, 105)
Another computer program, BACON.3 “…uses increasing heuristic levels to gather information, see if relations exist between new information and old, reject information that seems to have no effect on variables and telescope variables together that act the same on other variables.” But when all is said and done, Wagman concludes that “…BACON.3 …merely executed heuristic codes designed and interpreted by its developer.” (Wagman, 120, 126) It was still the human brain that provided the representation to the system, not to mention the interpretation both before and after the fact.
Much of the difficulty with our ability to define whether or not machines can be built that think revolves around our own definitions of thinking. Wagman defines fuzzy logic as a “….multivalued logic responsive to the imprecise character of human experience, judgement and reasoning.” (Wagman, 144) But there is evidence the human brain doesn’t operate by weighing each variable with a precise formula, but rather juggles variables according to past and ongoing present experience, and perhaps future presumptions, which will be different for each individual, and to satisfice those results rather than deciding on the best solution to a problem. For a human brain, fuzzy logic is truly fuzzy, that is, variables may not have particular values or weights, or may not be considered variables at all, but rather as fixed entities against a changing environment.
The limitations of the machine become evident at higher levels of complexity as well. As Waldrop suggests “…a knowledge base that mixes general-reasoning heuristics with domain knowledge has a way of turning concrete.” That is, “…it becomes next-to-impossible to add or modify rules without drastically - and unpredictably - changing the behaviour of the whole system.” (Waldrop, p. 42) The human brain, on the other hand, apparently operates more efficiently in problem solving with either more or less information. We are good at filling in the gaps, either with appropriate metaphors, or failing that, inappropriate ones. We are even able to find a large percentage of right answers with little information. Or we may come to wrong conclusions with a large database. The brain is apparently able to juggle such concepts effortlessly.
An excellent example of heuristics was provided at the SFU Conference on Cognition by Gird Girdsinger of MIT. He asked the question, “Which is bigger, San Antonio or San Diego?” of two groups of grad students at MIT and a German university. The MIT group, some of whom were geography students, got about 62% correct answers, the German group got 100% correct. Why? The German group used a simple heuristic that says roughly, “If you’ve heard of it, it’s probably bigger.” Girdsinger went on to design a series of stock market portfolios that showed consistently better success with the less-knowledgeable heuristics-using students than with the more knowledgable business-oriented students. (SFU XIth Conference on Cognition, 1998)
Shannon, like Chomsky, studied information apart from its semantic or meaning-carrying function, in order to tease out that part that can be analyzed accurately with empirical testing. For the study of mechanical intelligence, or the parameters of same, this has been particularly fruitful research. It is less clear that this method adds to our understanding of brain activity.
In 1975, Simon suggested to Pamela McCorduck “…four examples of networked computers…” (McCorduck, p. 76) found in nature, two of which may not have minds, and two of which he thinks do have minds: the solar system, weather system, economic system and the political system. While Shannon’s theorems are able to make reasonably accurate predictions within all four realms, there are problems in thinking of the latter two as computers. Economic and political systems may only seem to operate as machines, being a function of human activities. Human wants, for instance, which seem to be unlimited, play a large part in the functioning of both. The hypothesis is nevertheless open to empirical testing.
There are, according to Robert Baer, “…at least three important ways for things to go awry…” with computers. Either the “…computer is actually sick” (ie: dead transistor, etc.), there is “…a flaw in the operating system” or the trouble may be because of “…the finiteness of the computer.” (Baer, 84-85) These suggest limitations in the way a machine can process information that are different from the limitations of evolved brains. Economic and political systems may be particularly susceptible to the finiteness problem.
Shannon’s theorems suggest “Entropy is an aspect of probability, and probability, as deBeauregard asserts, ‘operates as the hinge between matter and mind…’” Further, “Probability is… a method of encoding partial knowledge… or missing information.” It “…measures both knowledge and ignorance, just as Shannon’s entropy does.” (Campbell, p. 33, 62, 65) The ability to ’satisfice’, to use heuristic reasoning to make best guesses and accept less-than-perfect answers to our questions indicates an evolved trait that is quite different from the abilities of machines, and may be impossible to duplicate in present day machines. Natural selection has probably operated to provide a means of dealing with problems that is markedly different from the problem-solving abilities we look to our machines to provide.
At the 1998 Symposium on Cognition at Simon Fraser, Deloros Dellarosa spoke on dominance hierarchies in a chimpanzee community. Because of the hierarchy, subdominant members of the group were able to survive not because of their ability to adapt to the constraints of their particular status, but rather their abilities to circumvent those constraints. The chimp who learns to hide her food from the more dominant members of her group is able to eat better.
Along with entropy and probability, in the understanding of brain and computer functions as information processing, is the concept of redundancy. Any complex system must, in order to communicate either within itself or to other systems, utilize some redundancy in order to form and transmit messages accurately in such a way that they can be understood by the receiver mechanism. “Redundancy is a means of keeping the system running in the presence of malfunction” and, “…redundancy makes complexity possible.” (Campbell, p. 73)
Genetic codes are basically simple but they carry a large amount of information which must be redundant enough to accurately define and codify a species from generation to generation. “The codes that made possible the higher organisms, Gatlin suggests, were redundant enough to ensure transmission along the channel from DNA to protein without error, yet at the same time they possessed an entropy, in Shannon’s sense of ‘amount of potential information’ high enough to generate a large variety of possible messages.” (Campbell, p. 113-114)
Gatlin proposes “…two different kinds of redundancy …[that] lower the entropy, but not in the same way.” D1, which is context free, measures the extent to which a sequence departs from the completely random… [and] D2, which is context sensitive, measures the extent to which the individual symbols have departed from a state of perfect independence from one another.” (Campbell, p. 118, 120) She hypothesizes that D2 redundancy in the DNA messages marks the evolution of the vertebrates from previous forms.
Redundancy is not as necessary in computer interactions as it is in evolved organisms, which may account for some of the differences in human thought and machine information processing. The former must transmit redundant messages which are inhibited and selected by satisficing and other means, whereas the latter is able to arrive at correct inferences with more direct and straightforward methods. Background noise is less a problem in a mechanical universe than in a chaotic one.
Campbell notes that “…as Shannon said, we are able to understand the brain better by studying the ways it is unlike a computer.” (Campbell, p. 199) Which is one method we might use to gain some understanding of where meaning comes from and how we are able to use our symbol-making and -manipulating abilities to develop a wide range of meanings from supposedly inanimate reality.
Many AI researchers believe the differences between human thought and what computers do when we say they are thinking may be only of magnitude. Others aren’t so sure. There are marked physical differences in the hardware of the different systems. The nerve axon is a complicated piece of work with a chemical signal that changes to an electrical, and then back to a chemical process as the synapse is bridged by the signal. The computer uses an entirely electrical process of transmission, much faster than chemical processes, but without the inhibitory processes that make brain activity able to handle redundancies and partial knowledge, among other abilities. Axon activities may not be reducible to binary notation as are computer chips, although some researchers insist everything is ultimately so reducible.
Regardless of the interpretation given, there are specific brain activities that have been studied fairly thoroughly, the vision process, memory, and the formation of language among others.
Tags: none
Category: essay |
