Copyright © 1999 Massachusetts Institute of Technology
No idea in the history of studying mind has fueled such enthusiasm -- or such criticism -- as the idea that the mind is a computer. This idea, known as "cognitivism" (Haugeland 1981/1998) or the COMPUTATIONAL THEORY OF MIND, claims not merely that our minds, like the weather, can be modeled on a computer, but more strongly that, at an appropriate level of abstraction, we are computers.
Whether cognitivism is true -- even what it means -- depends on what computation is. One strategy for answering that question is to defer to practice: to take as computational whatever society calls computational. Cognitivism's central tenet, in such a view, would be the thesis that people share with computers whatever constitutive or essential property binds computers into a coherent class. From such a vantage point, a theory of computation would be empirical, subject to experimental evidence. That is, a theory of computation would succeed or fail to the extent that it was able to account for the machines that made Silicon Valley famous: the devices we design, build, sell, use, and maintain.
Within cognitive science, however, cognitivism is usually understood in a more specific, theory laden, way: as building in one or more substantive claims about the nature of computing. Of these, the most influential (especially in cognitive science and artificial intelligence) has been the claim that computers are formal symbol manipulators (i.e., actively embodied FORMAL SYSTEMS). In this three-part characterization, the term "symbol" is taken to refer to any causally efficacious internal token of a concept, name, word, idea, representation, image, data structure, or other ingredient that represents or carries information about something else (see INTENTIONALITY). The predicate "formal" is understood in two simultaneous ways: positively, as something like shape, form, or syntax, and negatively, as independent of semantic properties, such as reference or truth. Manipulation refers to the fact that computation is an active, embodied process -- something that takes place in time. Together, they characterize computation as involving the active manipulation of semantic or intentional ingredients in a way that depends only on their formal properties. Given two data structures, for example, one encoding the fact that Socrates is a man, the other that all men are mortal, a computer (in this view) could draw the conclusion that Socrates is mortal formally (i.e., would honor what Fodor 1975 calls the "formality condition") if and only if it could do so by reacting to the form or shape of the implicated data structures, without regard to their truth, reference, or MEANING.
In spite of its historical importance, however, it turns out that the formal symbol manipulation construal of computing is only one of half a dozen different ideas at play in present-day intellectual discourse.
Among alternatives, most significant is the body of work carried on in theoretical computer science under the label theory of computation. This largely automata-theoretic conception is based on Turing's famous construction of a "Turing machine": a finite state controller able to read, write, and move sequentially back and forth along an infinitely long tape that is inscribed with a finite but indefinite number of tokens or marks. It is this Turing-based theory that explains the notion of a universal computer on which the CHURCH-TURING THESIS rests (that anything that can be done algorithmically at all can be done by a computer). Although the architecture of Turing machines is specific, Turing's original paper (1936/1965) introduced the now ubiquitous idea of implementing one (virtual) architecture on top of another, and so the Church-Turing thesis is not considered to be architecturally bound. The formal theory of Turing machines is relatively abstract, and used primarily for theoretical purposes: to show what can and cannot be computed, to demonstrate complexity results (how much work is required to solve certain classes of problems), to assign formal semantics to programming languages, etc. Turing machines have also figured in cognitive science at a more imaginative level, for example in the classic formulation of the Turing test: a proposal that a computer can be counted as intelligent if it is able to mimic a person answering a series of typed questions sufficiently well to fool an observer.
Because of their prominence in theoretical computer science, Turing machines are often thought to capture the essential nature of computing -- even though, interestingly, they are not explicitly characterized in terms of either SYNTAX or formality. Nor do these two models, formal systems and Turing machines, exhaust extant ideas about the fundamental nature of computing. One popular third alternative is that computation is information processing -- a broad intuition that can be broken down into a variety of subreadings, including syntactic or quantitative variants (à la Shannon, Kolmogorov, Chaitin, etc.), semantic versions (à la Dretske, Barwise and Perry, Halpern, Rosenschein, etc.), and historical or socio-technical readings, such as the notion of information that undergirds public conceptions of the Internet. Yet another idea is that the essence of being a computer is to be a digital state machine. In spite of various proofs of broad behavioral equivalence among these proposals, for purposes of cognitive science it is essential to recognize that these accounts are conceptually distinct. The notion of formal symbol manipulation makes implicit reference to both syntax and semantics, for example, whereas the notion of a digital machine does neither; the theory of Turing computability is primarily concerned with input-output behavior, whereas formal symbol manipulation focuses on internal mechanisms and ways of working. The proposals are distinct in extension as well, as continuous Turing machines can solve problems unsolvable by digital (discrete) computers; not all digital machines (such as Lincoln Log houses) are symbol manipulators; and information, especially in its semantic reading as a counterfactual- supporting correlation, seems unrestricted by the causal locality that constrains the notion of effectiveness implicit in Turing machines.
It is not only at the level of basic theory or model that computation affects cognitive science. Whatever their underlying theoretical status, it is undeniable that computers have had an unprecedented effect on the human lives that cognitive science studies -- both technological (e-mail, electronic commerce, virtual reality, on-line communities, etc.) and intellectual (in artificial intelligence projects, attempts to fuse computation and quantum mechanics, and the like). It is not yet clear, however, how these higher level developments relate to technical debates within cognitive science itself. By far the majority of real-world systems are written in such high-level programming languages as Fortran, C++, and Java, for example. The total variety of software architectures built in such languages is staggering -- including data and control structures of seemingly unlimited variety (parallel and serial, local and distributed, declarative and procedural, sentential and imagistic, etc.). To date, however, this architectural profusion has not been fully recognized within cognitive science. Internal to cognitive science, the label "computational" is often associated with one rather specific class of architectures: serial systems based on relatively fixed, symbolic, explicit, discrete, high-level representations, exemplified by axiomatic inference systems, KNOWLEDGE REPRESENTATION systems, KNOWLEDGE-BASED SYSTEMS, etc. (van Gelder 1995).
This classic symbolic architecture is defended, for example by Fodor and Pylyshyn (1988), because of its claimed superiority in dealing with the systematicity, productivity, and compositionality of high-level thought. Its very specificity, however, especially in contrast with the wide variety of architectures increasingly deployed in computational practice, seems responsible for a variety of self-described "non-" or even "anti-" computational movements that have sprung up in cognitive science over the last two decades: connectionist or neurally inspired architectures (see COGNITIVE MODELING, CONNECTIONIST and CONNECTIONISM, PHILOSOPHICAL ISSUES); DYNAMIC APPROACHES TO COGNITION; shifts in emphasis from computational models to cognitive neuroscience; embodied and situated architectures that exploit various forms of environmental context dependence (see SITUATEDNESS/ EMBEDDEDNESS); so-called complex adaptive systems and models of ARTIFICIAL LIFE, motivated in part by such biological phenomena as random mutation and evolutionary selection; interactive or even merely reactive robotics, in the style of Brooks (1997), Agre and Chapman (1987); and other movements. These alternatives differ from the classical model along a number of dimensions: (i) they are more likely to be parallel than serial; (ii) their ingredient structures, particularly at the design stage, are unlikely to be assigned representational or semantic content; (iii) such content or interpretation as is ultimately assigned is typically based on use or experience rather than pure denotation or description, and to be "nonconceptual" in flavor, rather than "conceptualist" or "symbolic"; (iv) they are more likely to be used to model action, perception, navigation, motor control, bodily movement, and other forms of coupled engagement with the world, instead of the traditional emphasis on detached and even deductive ratiocination; (v) they are much more likely to be "real time," in the sense of requiring a close coupling between the temporality of the computational process and the temporality of the subject domain; and (vi) a higher priority is typically accorded to flexibility and the ability to cope with unexpected richness and environmental variation rather than to deep reasoning or purely deductive skills.
Taken collectively, these changes represent a sea change in theoretical orientation within cognitive science -- often described in contrast with traditional "computational" approaches. Without a richer and more widely agreed upon theory of computation, however, able to account for the variety of real-world computing, it is unclear that these newly proposed systems should be counted as genuinely "noncomputational." No one would argue that these new proposals escape the COMPUTATIONAL COMPLEXITY limits established by computer science's core mathematical theory, for example. Nor can one deny that real-world computing systems are moving quickly in many of the same directions (toward parallelism, real-time interaction, etc.). Indeed, the evolutionary pace of real-world computing is so relentless that it seems a safe bet to suppose that, overall, society's understanding of "computation" will adapt, as necessary, to subsume all these recent developments within its ever-expanding scope.
Agre, P., and D. Chapman. (1987). Pengi: An implementation of a theory of activity. In The Proceedings of the Sixth National Conference on Artificial Intelligence. American Association for Artificial Intelligence. Seattle: Morgan Kaufmann, pp. 268-272.
Barwise, J., and J. Perry. (1983). Situations and Attitudes. Cambridge, MA: MIT Press.
Brooks, R. A. (1997). Intelligence without representation. Artificial Intelligence 47:139-159. In J. Haugeland, Ed., Mind Design II: Philosophy, Psychology, Artificial Intelligence. 2nd ed., revised and enlarged. Cambridge, MA: MIT Press, pp. 395-420.
Dretske, F. (1981). Knowledge and the Flow of Information. Cambridge, MA: MIT Press.
Fodor, J. (1975). The Language of Thought. Cambridge, MA: MIT Press.
Fodor, J., and Z. Pylyshyn. (1988). Connectionism and cognitive architecture: A critical analysis. In S. Pinker and J. Mehler, Eds., Connections and Symbols. Cambridge, MA: MIT Press, pp. 3-71.
Haugeland, J. (1981/1998). The nature and plausibility of cognitivism. Behavioral and Brain Sciences I: 215-226. Reprinted in J. Haugeland, Ed., Having Thought: Essays in the Metaphysics of Mind. Cambridge, MA: Harvard University Press, 1988, pp. 9 - 45.
Turing, A. M. (1936/1965). On computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society 2nd ser. 42:230-265. Reprinted in M. Davis, Ed., The Undecidable: Basic Papers on Undecidable Propositions, Unsolvable Problems and Computable Functions. Hewlett, NY: Raven Press, 1965, pp. 116 - 154.
van Gelder, T. J. (1995). What might cognition be, if not computation? Journal of Philosophy 91:345-381.
Copyright © 1999 Massachusetts Institute of Technology