---

 

I. FROM COMPUTERS TO MECHANICAL MINDS?
II. SYMBOLIC SYSTEMS-2: BRAIN MODELING-1
III. THE RETURN OF THE BRAIN MODELING PARADIGM

  "What a man sees depends both upon what he looks at and also upon what his previous visual-conceptual experience has taught him to see." Thomas S. Kuhn
 

 Following Malcolm Spector's and John Kitsuse's definition of social problems as claims-making activities, this paper examines the claims-making activities in artificial intelligence research from a social problems constructionist perspective.  The constructionist perspective views social problems as constructed through the claims-making activities of interested groups.  "The activity of making claims, complaints, or demands for change is the core of what we call social problems activities.  Definitions of conditions as social problems are constructed by members of a society who attempt to call attention to situations they find repugnant and who try to mobilize the institutions to do something about them."

Artificial intelligence (AI) research, the search for mechanical thinking ability, has been the subject of scientific and philosophical debate and claims-making activities within the academic/scientific community beginning with the inception of the idea .  Thomas S. Kuhn's descriptions of scientific revolutions based in scientific paradigms (world-views) provides structure for examining the history of AI research claims-making.  Kuhn claimed that science is not cumulative as it is presented in history texts, but that it is structured in group viewpoints of the scientists.  These group viewpoints he labeled as paradigms.  A scientific paradigm structures the kinds of questions a scientist can ask when doing "normal science."  For a new scientific paradigm to replace an older paradigm requires a "Gestalt" experience for the scientists .   As a social problem the AI debate has led to claims on all sides of the issue far exceeding the accomplishments of the claimants; and the channeling of public research funds in directions based on these claims .  The primary reason for this claims-making has been the specific paradigm or world view that the AI researchers have held.

 This paper is divided into three sections describing the claims-making activities of AI researchers.  The first outlines the history of AI research and the emergence of two paradigms of AI research.  The second describes the results of the conflict between the two research paradigms; the philosophically based AI research and the brain modeling based AI research.  The third summarizes and describes current claims-making activities in this field of research.                                                         <-top

I. FROM COMPUTERS TO MECHANICAL MINDS?
 

 "Ask a dozen different researchers the question 'What is AI?' and you get a dozen different answers.  Such is not the case with more mature disciplines, such as physics, medicine, and chemistry.  To some people, this is evidence that AI can't be classified as a science and that it is, rather, simply a software-engineering discipline that has taken on airs.  But this view discounts the fact that every mature science was once immature and groping for definition."
          Bob Ryan.  "AI's Identity Crisis" in Byte.
          Jan. 1991. Page 239.
 
 The first electronic digital computers were built in the late forties replacing in most applications the larger and slower analog experiments.  Within a few years, by the early fifties, the quest to turn computers into machines that could duplicate the human thinking process emerged with two separate paradigms or factions.  One paradigm viewed human thought as a mechanical/chemical function of the human brain and sought to duplicate the neural networks of the brain.  This paradigm descended from "idealized, holistic neuroscience"  and became known as the "bottom up" approach because it started with the physical system of the "brain" as its model .  The other faction believed that human thought consisted of symbol manipulation and that a machine could be built to think if it was given the right basic symbols .  This paradigm descended from a "rationalist, reductionist" philosophical viewpoint  and became known as the "top down" approach because focused on symbolic representations manipulated by the "mind" as the model of the way human's think .

 Part of the problem for the sociologist or layman in identifying the claims-making activities present in the artificial intelligence research field is based on the normal proliferation of jargon in any specialized endeavor that is used to describe that specialized endeavor.  Many terms are used within the AI research field to describe the same two basic paradigmatic research bases.  Some examples are given in the following paragraphs.  These two research paradigms developed at approximately the same time, but with very different research focuses.

 One scientific viewpoint was grounded in physical systems.  They believed thinking was a product of the complicated chemical machine, the human brain.  This neuroscience (brain simulation) based paradigm became known as the "bottom-up" approach.  Scientists working with this approach viewed computer hardware as the place to focus their research.  Jargon for mechanical processes designed to simulate brain function include such terms as "cybernetics," "perceptrons," "neural modeling," "brain modeling," "neural nets,"  "multi-layer machines," "materialists"  and "connectionist" .
 The other model was grounded in language as a symbolic representation of the world.  These scientists believed thought processes could be reduced to basic symbols, and that these basic symbols could then be programmed into a computer to teach it to think.  They drew their view from certain philosophical explorations such as Wittgenstein's Tractatus.  Bertrand Russell, Wittgenstein, Descartes and Goedel were philosophers engaging in philosophical phenomenology - the search for atomic facts and basic objects.  Unfortunately these AI researchers seemed unaware of Wittgenstein's later recant of his earlier writings in  Philosophical Investigations, published in 1953 , or Goedel's mathematical "Incompleteness Theorem" .  "Goedel's theorem states that in any consistent system which is strong enough to produce simple arithmetic, there are formulae which cannot be proved in the system but which we can see to be true" .  Because of this channeling toward philosophically based research grounded in the idea of basic symbols, little progress was made in AI research in the early 1980's.

This approach to AI research became known as the "top-down" approach and worked with computer programs (software) trying to duplicate human thought processes.  Terms used for this type of research include "complex information processing," "dualists," "symbolic manipulation," and "symbolic systems." One source of the neuroscience based AI paradigm's beginnings can be identified with the book, Cybernetics, written by Massachusetts Institute of Technology (M.I.T.) professor Norbert Wiener and first published in 1948.  He argued that feedback was the way creatures, including human beings, learn about and adapt to their environment .  "This cybernetic, or neural-modeling, approach to machine intelligence was soon dubbed the 'bottom-up' approach" with the goal of starting with a model of the brain function in the primitive organism and working up to a human equivalent .   The silicon computer chip was yet to be invented, and hardware limitations proved expensively daunting to many early researchers.

 Frank Rosenblatt, a research psychologist at the Cornell Aeronautical Laboratory optimistically continued this line of research, and in 1958 staged a demonstration of his "perceptron," an IBM 704 computer connected to an "eye" made of photoelectric cells and programmed to distinguish between two patterns of squares . Rosenblatt figured that it "is both easier and more profitable to axiomatize the physical system and then investigate this system analytically to determine its behavior, than to axiomatize the behavior and then design a physical system by techniques of logical synthesis" .  In 1960, one year beyond his own projected deadline, Rosenblatt demonstrated the Mark I, a perceptron that could learn to make slight discriminations between letters of the alphabet through a "trial-and-error" process.

 "Rosenblatt predicted a bright future for the Perceptron.  He thought that it would be able to teach itself.  He even thought that the Perceptron would eventually be able to build other Perceptrons" .
The Mark I lacked flexibility, being unable to recognize partial letters, or letters facing unusual directions .  "This problem is a mystery that still baffles AI workers and computer scientists.  How can a computer be taught when to break the rules it has been programmed to obey ?"

  In 1956, Allen Newell, Herbert A. Simon, and J. C. Shaw  preferred to call their research "complex information processing" .  They advocated a "top-down" approach to AI research primarily because software could be easily modified, and failure more easily abandoned .  They created a computer program called Logic Theorist using the philosophical treatise on mathematics, Principia Mathematica, by Alfred North Whitehead and Bertrand Russell as their inspiration.  Their primary programming task was to avoid what AI researchers call "combinatorial explosion;" as the number of variables considered increases, the number of combinations capable of being created increases exponentially .  As yet, AI researchers have been unable to cope with the infinite diversity - infinite combination capable of being processed by the human mind .
 At this point, Simon predicted that through AI research "digital computers would be the world's chess champions" and "discover at least one important new mathematical theorem" within 10 years .  They followed Logic Theorist with an improved version in 1957 that they called General Problem Solver (GPS) capable of a limited heuristic problem solving method .  A heuristic problem solving method starts with a goal and makes choices appearing to approach that goal.  An example of this would be when trying to get to the downtown area in an unfamiliar city, we keep turning on to the streets that lead in the direction of the large buildings located downtown.

 Encouraged by his program General Problem Solver (GPS), Herbert Simon made the claim in 1957 that

 "...there are now in the world machines that think, that learn and that create.  Moreover, their ability to do these things is going to increase rapidly until- in a visible future- the range of problems they can handle will be coextensive with the range to which the human mind has been applied.".
 Writing of these two AI research paradigms, University of California at Berkeley professors, Hubert L. Dreyfus and Stuart E. Dreyfus commented:
 "In the early sixties both approaches looked equally promising, and both made themselves equally vulnerable by making exaggerated claims.  Yet the results of the internal war between the two research programs were surprisingly asymmetrical.  By 1970 the brain simulation research, which had its paradigm in the perceptron, was reduced to a few lonely, underfunded efforts, while those who proposed using digital computers as symbol manipulators had undisputed control of the resources, graduate programs, journals and symposia that constitute a flourishing research program.".
Why, at this early stage, did one paradigm, without clearly addressing many of the questions being asked, triumph over the other?  How did the symbolic systems paradigm gain control of the research funds while the brain modeling paradigm took a back seat?  The next section of this paper examines explanations of how this happened.    <-top

II. SYMBOLIC SYSTEMS-2: BRAIN MODELING-1
 

 "AI is internally in a paradigmatic mess.  There is really no broad agreement on the essential nature or formal basis of intelligence and the proper theoretical framework for it."  B. Chandrasekaran of Ohio State
 Thomas Kuhn thought that a new scientific research paradigm would replace a competing older scientific paradigm when sufficient anomalies were explained by the newer model .  But two new paradigms of a new science competed for AI research funds.  In less than twenty years of research one paradigm of research almost totally obscured the other, but not by answering research questions since neither had achieved their exaggerated claims.  Each group had both advocates and detractors.  The criticism of the detractors of each group was basically the same: they could solve certain limited problems, but once the complexity of the real world is encountered, their methods bog-down .  One group of advocates within the symbolic systems paradigm chose to specifically criticize the brain modeling paradigm.

 Two AI researchers are credited with the supersedence of the symbolic systems model over the neuroscience approach.  They succeeded in channeling the majority of research funds in the direction of the "top-down"  symbolic based research .  High school classmate to Rosenblatt and professor at M.I.T., Marvin Minsky met Seymour Papert, another M.I.T. professor, and together they became advocates of the perspective known as the "top-down" approach .  In 1965, Minsky and Papert began circulating through the field of AI researchers a draft of their book, Perceptrons, attacking Rosenblatt's perceptron work.  In their book they described writings about the perceptron research as being "without scientific value" .  Minsky and Papert clearly attack the brain-modeling research as a paradigm (as defined by Thomas Kuhn) when they go on to write:

 "Both of the present authors (first independently and later together) became involved with a somewhat therapeutic compulsion: to dispel what we feared to be the first shadows of a 'holistic' or 'Gestalt' misconception that would threaten to haunt the fields of engineering and artificial intelligence as it had earlier haunted biology and psychology."
 Hubert L. Dreyfus and Stuart E. Dreyfus claim that "Minsky and Papert were so intent on eliminating all competition" and so entrenched in their symbolic research paradigm, that they failed to examine "multilayer," also known as "neural net"  or "connectionist" machines .
 "Their attack on gestalt thinking in AI succeeded beyond their wildest dreams....But why was it enough? It was too early to close accounts on either approach.  Yet something in Minsky and Papert's book struck a responsive chord.  It seemed AI workers shared the quasi-religious philosophical prejudice against holism that motivated the attack."
 
Given this philosophical prejudice of some of the workers in AI research, the limited physical computer systems available at the time and the costs of these systems, the brain modeling researchers could do little more than speculate about the potential of their approach, while symbolic systems researchers were creating programs with some usefulness.                                  <-top

III. THE RETURN OF THE BRAIN MODELING PARADIGM
 

 "Neither proof, nor error is at issue, the transfer of allegiance from paradigm to paradigm is a conversion experience that cannot be forced."  Thomas S. Kuhn
 Although research funds had been diverted from the brain modeling research paradigm and many AI researcher thought that line of research was a dead-end, a few AI researchers continued in this line.  The advent of cheap computer memory combined with the invention of the silicon computer chip gave new life to the hardware based brain modeling paradigm.  By the mid eighties, brain modeling based research was back in competition with the symbolic researchers for research funds .  Of this turn around, Bob Ryan wrote for Byte magazine:
 "As connectionist machines continue to produce results in areas such as speech recognition and machine vision -- areas formerly the exclusive reserve of symbolic AI-- the debate about which method is best for representing and processing knowledge will intensify.... Undoubtedly, as old antagonisms wear thin, there will be more research into systems that combine both approaches."
 If philosopher Ludwig Wittgenstein is right and human thought is a much more holistic process than either neural nets or symbol systems; then the future paradigm of AI research will be a combination of both paradigms rather than one paradigm replacing the other as Thomas Kuhn's work would indicate.

 Current successful AI research is certainly headed in the direction of paradigm synthesis.  The most successful symbolic based research outcome, so far, has been "expert systems," computer programs that locate data in a specific area of expertise such as medicine, law, or a specific business application .  AI researchers combining expert systems and neural nets are creating hybrid systems with the ability to provide feedback to improve each other.  Don Barker, coordinator of the Computer Assisted Learning Center at Gonzaga University, says

 " symbolic (expert systems) and parallel distributed processing (neural networks) are not competing AI strategies but complementary.  By uniting them, we can avoid many of the weaknesses inherent in each method while capitalizing on their unique strengths."
AI researchers may not ever duplicate human thought and intelligence with machines, but the results of their research has proved useful so far .  We will see what the future brings.
 

   Top of page       Home