Connectionism, Philosophical Issues

Since its inception, artificial intelligence (AI) research has had a growing influence on the philosophy of mind. Consequently, the recent development of a radically different style of cognitive modeling -- commonly known as "connectionism" (see COGNITIVE MODELING, CONNECTIONIST) -- has brought with it a number of important philosophical issues and concerns. Because connectionism is such a dramatic departure from more traditional accounts of cognition, it has forced philosophers to reconsider several assumptions based on earlier theories. Most of these cluster around three central themes: (1) the nature of psychological explanation, (2) forms of mental representation, and (3) nativist and empiricist accounts of learning.

Before the introduction of connectionism in the mid-1980s, the dominant paradigm in cognitive modeling was the COMPUTATIONAL THEORY OF MIND, sometimes referred to by philosophers as "GOFAI" (for "good old-fashioned artificial intelligence"; see also COGNITIVE MODELING, SYMBOLIC). GOFAI accounts treat the mind as a complex organization of interacting subsystems, each performing a specific cognitive function and processing information through the manipulation of discrete, quasi-linguistic symbols whose interactions are governed by explicitly encoded rules. Psychological explanation is treated as a form of FUNCTIONAL DECOMPOSITION, where sophisticated cognitive capacities are broken down and explained through the coordinated activity of individual components. The capacities of the individual components are further explained through a description of their internal symbolic operations (Cummins 1983; see also RULES AND REPRESENTATIONS and ALGORITHM).

Connectionism suggests a very different outlook on the nature of psychological theory. Connectionist networks model cognition through the spreading activation of numerous simple units. The processing is highly distributed throughout the entire system, and there are no task-specific modules, discrete symbols, or explicit rules that govern the operations (Rumelhart, McClelland, and PDP Research Group 1986; McClelland, Rumelhart, and PDP Research Group 1986; Smolensky 1988). This has forced researchers to abandon the functional decomposition approach and search for new ways to understand the structure of psychological explanation. In one popular alternative, DYNAMIC APPROACHES TO COGNITION, cognitive activity is understood as a series of mathematical state transitions plotted along different possible trajectories. Mental operations are described through equations that capture the behavior of the whole system, rather than focusing on the logical or syntactic transformations within specific subsystems. Some writers believe this framework will provide a new paradigm for understanding the nature of COGNITIVE ARCHITECTURE and give rise to psychological explanations that depart dramatically from past accounts (van Gelder 1991; Horgan and Tienson 1996; see also COMPUTATIONAL NEUROSCIENCE).

GOFAI cognitive models rely heavily on explicit, syntactically structured symbols to store and process information. By contrast, connectionist networks employ a very different type of representation, whereby information is encoded throughout the nodes and connections of the entire network. These distributed representations (cf. DISTRIBUTED VS. LOCAL REPRESENTATION) lack the languagelike, syntactic structure of traditional GOFAI symbols. Moreover, their content and representational function is often revealed only through mathematical analysis of the activity patterns of the system's internal units.

The philosophical implications of this new account of representation are far-reaching. Some writers, unhappy with the quasi-linguistic character of GOFAI symbols, have embraced the connectionist picture to support nonsentential theories of representation, including prototype accounts of CONCEPTS (Churchland 1989). Others have suggested parallels between the connectionist representations and the biologically motivated theories of INFORMATIONAL SEMANTICS explored by writers such as Fred Dretske (1988). Many believe the internal units of connectionist networks provide a promising new way to understand MENTAL REPRESENTATION because of their similarity to real neural systems and their sensitivity to environmental stimuli (Bechtel 1989).

On the other hand, some philosophers have argued that the connectionist account of representation is seriously flawed. Jerry Fodor and Zenon Pylyshyn have claimed that the ability to represent some states of affairs (e.g., "John loves Mary") is closely linked to the ability to represent other states of affairs (e.g., "Mary loves John"; Fodor and Pylyshyn 1988). They argue that this feature of cognition, called "systematicity," must be explained by any plausible theory of mind. They insist that because connectionist representations do not have constituent parts, connectionist models cannot explain systematicity. In response to this challenge, several connectionists have argued that it is possible for connectionist representations to produce systematic cognition in subtle ways without merely implementing a symbolic system (Smolensky 1991; Clark 1991).

Connectionist accounts of representations have also influenced philosophical debate concerning the status of PROPOSITIONAL ATTITUDES. ELIMINATIVE MATERIALISM holds that our commonsense conception of the mind is so flawed that there is reason to be skeptical about the existence of states such as beliefs and desires. Some writers have suggested that the style of information encoding in networks is so radically different from what is assumed by common sense that connectionist models actually give credence to eliminativism (Churchland 1986; Ramsey, Stich, and Garon 1990). Others have gone a step further and argued that the internal elements of networks should not be viewed as representations at all (Brooks 1991; Ramsey 1997). In res-ponse, several writers have insisted that commonsense psychology and connectionism are quite compatible, once the former is properly construed; moreover, because our commonsense notion of belief is not committed to any specific sort of cognitive architecture, it has nothing to fear from the success of connectionism (Dennett 1991; Bechtel and Abrahamsen 1993; see also FOLK PSYCHOLOGY).

Research in cognitive science has had an important influence on the traditional debate between nativists, who claim that we are born with innate knowledge, and empiricists, who claim that knowledge is derived from experience (see also NATIVISM and RATIONALISM VS. EMPIRICISM). Nativism has enjoyed popularity in cognitive science because it has proven difficult to explain how cognitive capacities are acquired without assuming some form of preexisting knowledge within the system. Yet one of the most striking features of connectionist networks is their ability to attain capacities with very little help from antecedent knowledge. By relying on environmental stimuli and powerful learning algorithms, networks often appear to program themselves. This has led many to claim that connectionism offers a powerful new approach to learning -- one that will resurrect empiricist accounts of the mind.

A mainspring of nativism in cognitive science has been Chomsky's POVERTY OF THE STIMULUS ARGUMENT for the INNATENESS OF LANGUAGE (1975). Chomsky has argued that LANGUAGE ACQUISITION is impossible without a rich store of innate linguistic knowledge. Although several CONNECTIONIST APPROACHES TO LANGUAGE have been developed to demonstrate how areas of linguistic competence -- such as knowing regular and irregular past tense forms of verbs -- can be obtained without preexisting linguistic rules (Rumelhart, McClelland, and PDP Research Group 1986; Elman et al. 1996), the success of these models in establishing a nonnativist theory of linguistic competence has been heavily debated (Pinker and Prince 1988). One critical issue concerns the degree of DOMAIN SPECIFICITY employed in the learning strategies and initial configuration of the networks (Ramsey and Stich 1990).

A second motivation for nativism stems from the "classical" account of concept acquisition, which assumes that learning occurs when new complex concepts are constructed from more primitive concepts (Fodor 1981), and which suggests there must first exist a prior store of basic concepts that, by hypothesis, are unlearned. However, connectionism appears to offer a different model of concept acquisition. Networks seem to develop new classifications and abstractions that emerge without the recombination of preexisting representations. In other words, there is reason to think connectionist learning gives rise to new primitive concepts that are developed entirely in response to the system's training input (Munakata et al. 1997). To many, this captures the essence of empiricist learning and signals a new direction for understanding CONCEPTUAL CHANGE (Churchland 1989; Elman et al. 1996).

Additional links

-- William Ramsey


Bechtel, W. (1989). Connectionism and intentionality. In Proceedings of the Eleventh Annual Meetings of the Cognitive Science Society. Hillsdale, NJ: Erlbaum, pp. 553-600.

Bechtel, W., and A. Abrahamsen. (1993). Connectionism and the future of folk psychology. In S. Christensen and D. Turner, Eds., Folk Psychology and the Philosophy of Mind. Hillsdale, NJ: Erlbaum, pp. 340-367.

Brooks, R. (1991). Intelligence without representation. Artificial Intelligence 47:139-159.

Chomsky, N. (1975). Reflections on Language. New York: Pantheon.

Churchland, P. M. (1986). Some reductive strategies in cognitive neurobiology. Mind 95(379):279-309.

Churchland, P. M. (1989). A Neurocomputational Perspective. Cambridge, MA: MIT Press.

Clark, A. (1991). Systematicity, structured representations and cognitive architecture: A reply to Fodor and Pylyshyn. In T. Horgan and J. Tienson, Eds., Connectionism and the Philosophy of Mind. Dordrecht: Kluwer, pp. 198-218.

Cummins, R. (1983). The Nature of Psychological Explanation. Cambridge, MA: MIT Press.

Dennett, D. (1991). Two contrasts: Folk craft versus folk science, and belief vs. opinion. In J. Greenwood, Ed., The Future of Folk Psychology. New York: Cambridge University Press, pp. 135-148.

Dretske, F. (1988). Explaining Behavior. Cambridge, MA: MIT Press.

Elman, J., E. Bates, M. Johnson, A. Karmiloff-Smith, D. Parisi, and K. Plunkett. (1996). Rethinking Innateness: A Connectionist Perspective on Development. Cambridge, MA: MIT Press.

Fodor, J. (1981). The present status of the innateness controversy. In Representations. Cambridge, MA: MIT Press, pp. 257-316.

Fodor, J., and Z. Pylyshyn. (1988). Connectionism and cognitive architecture: A critical analysis. Cognition 28:3-71.

Horgan, T., and J. Tienson. (1996). Connectionism and the Philosophy of Psychology. Cambridge, MA: MIT Press.

McClelland, J., D. Rumelhart, and PDP Research Group. (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol. 2, Psychological and Biological Models. Cambridge, MA: MIT Press.

Muntakata, Y., J. L. McClelland, M. H. Johnson, and R. S. Siegler. (1997). Rethinking infant knowledge: Toward an adaptive process account of success and failure in object permanence tasks. Psychological Review 104(4):686-713.

Pinker, S., and A. Prince. (1988). On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition 28:73-193.

Ramsey, W. (1997). Do connectionist representations earn their explanatory keep? Mind and Language 12(1):34-66.

Ramsey, W., and S. Stich (1990). Connectionism and three levels of nativism. Synthèse 82:177-205.

Ramsey, W., S. Stich, and J. Garon (1990). Connectionism, eliminativism and the future of folk psychology. Philosophical Perspectives 4:499-533.

Rumelhart, D., J. McClelland, and PDP Research Group. (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol. 1, Foundations. Cambridge, MA: MIT Press.

Smolensky, P. (1988). On the proper treatment of connectionism. Behavioral and Brain Sciences 11(1):1-74.

Smolensky, P. (1991). The constituent structure of mental states: A reply to Fodor and Pylyshyn. In T. Horgan and J. Tienson, Eds., Connectionism and the Philosophy of Mind. Dordrecht: Kluwer, pp. 281-308.

van Gelder, T. (1991). Connectionism and dynamical explanation. In Proceedings of the Thirteenth Annual Conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum, pp. 499-503.

Further Readings

Aizawa, K. (1994). Representation without rules, connectionism and the syntactic argument. Synthèse 101:465-492.

Bechtel, W. (1991). Connectionism and the philosophy of mind: An overview. In T. Horgan and J. Tienson, Eds., Connectionism and the Philosophy of Mind. Dordrecht: Kluwer, pp. 30-59.

Bechtel, W., and A. Abrahamsen. (1991). Connectionism and the Mind: An Introduction to Parallel Processing in Networks. Oxford: Blackwell.

Bechtel, W., and R. Richardson. (1993). Discovering Complexity. Princeton, NJ: Princeton University.

Chalmers, D. (1993). Connectionism and compositionality: Why Fodor and Pylyshyn were wrong. Philosophical Psychology 6:305-319.

Churchland, P. S., and T. Sejnowski. (1992). The Computational Brain. Cambridge, MA: MIT Press.

Clark, A. (1989). Microcognition. Cambridge MA: MIT Press.

Clark, A. (1993). Associative Engines: Connectionism, Concepts and Representational Change. Cambridge MA: MIT Press.

Forster, M. R., and E. Saidel. (1994). Connectionism and the fate of folk psychology: A reply to Ramsey, Stich and Garon. Philosophical Psychology 7:437-452.

Garson, J. (1994). Cognition without classical architecture. Synthèse 100:291-306.

Hanson, S., and D. Burr. (1990). What connectionist models learn: Learning and representation in connectionist networks. Behavioral and Brain Sciences 13:471-518.

Haugeland, J. (1978). The nature and plausibility of cognitivism. Behavioral and Brain Sciences 2:215-260.

Horgan, T., and J. Tienson, Eds. (1991). Connectionism and the Philosophy of Mind. Dordrecht: Kluwer.

Lloyd, D. (1989). Simple Minds. Cambridge, MA: MIT Press.

Nadel, L., L. Cooper, P. Culicover, and R. M. Harnish, Eds. (1989). Neural Connections, Mental Computation. Cambridge, MA: MIT Press.

Macdonald, C., and G. Macdonald, Eds. (1995). Connectionism: Debates on Psychological Explanation. Oxford: Blackwell.

McLaughlin, B. (1993). The connectionism/classicism battle to win souls. Philosophical Studies 71:163-190.

McLaughlin, B., and T. Warfield. (1994). The allure of connectionism re-examined. Synthèse 101:365-400.

Port, F., and T. van Gelder, Eds. (1995). Mind as Motion: Explorations in the Dynamics of Cognition. Cambridge, MA: MIT Press.

Ramsey, W., S. Stich, and D. Rumelhart, Eds. (1991). Philosophy and Connectionist Theory. Hillsdale, NJ: Erlbaum.

Tomberlin, J., Ed. (1995). Philosophical Perspectives. Vol. 9, AI, Connectionism and the Philosophy of Mind. Atascadero, CA: Ridgeview.