Young (1998) argues that my critique of connectionism (Green 1998) is grounded in an assumption that realism is superior to instrumentalism as an interpretation of scientific theories, and that the difficulties that I argue connectionism faces can be avoided if connectionists adopt an instrumentalist stance. I made no such assumption, however, and a closer examination of instrumentalism shows it to be detrimental to the connectionist cause. Realism -- probably neural realism -- remains the connectionist's best hope.
2. It appears that Young would have us return to the positivism (of a sort) of our behaviorist forebears, averring that our theories are not meant to represent the world in any realistic way, but merely to serve as "inference tickets" to the observations of which we want accounts. A lot of water has flowed beneath the philosophy-of-science bridge since the days of MacCorquodale and Meehl (1948). Instrumentalism is under attack from many quarters, mainly because it fails to provide any real EXPLANATION for the phenomena under study (see Van Fraassen 1980 for the most influential contemporary attempt to defend an "anti-realist" position while evading the standard realist criticisms of instrumentalism).
3. Nevertheless, Young raises a very important question: can connectionists evade my critique if they adopt an instrumentalist interpretation of their work? Even if the current criticisms of instrumentalism were to be overcome, I am not convinced that an instrumentalist interpretation of connectionist models of cognition would do the trick. The reasons are contained within Young's own commentary. He cites Skinner's (1950) recognition of the "need for a formal representation of the data reduced to a minimal number of terms" (cited in Young 1998, para. 3). The question is whether connectionist networks really do reduce it to a "minimal number of terms."
4. As Young himself writes: "One must be careful, however, to minimize the variables included in a theory. Introducing too many variables increases the theory's degrees of freedom, making it more likely that the theory will overfit the currently available empirical data. This overfitting will undermine the future validity of the theory" (para. 6). If one is a true instrumentalist, one is committed to NOTHING BUT accounting for the largest amount of data with the smallest number of theoretical terms. Because symbolic models, as I argue, almost invariably use far fewer theoretical terms than connectionist ones (often whole orders of magnitude fewer), the symbolists will almost always win such a fight (assuming they can account for an amount of data even moderately close to that accounted for by their connectionist counterparts). In other words, parsimony is anything BUT the connectionist's friend.
5. By contrast, realism relegates parsimony to a secondary role, appealing instead to factors not found in the Spartan principle of the instrumentalist, namely: does the theory actually represent what is going on "out there" in the world, whether parsimonious or not? Thus, parsimony can be invoked DEFENSIVELY by the connectionist, who can claim that although the symbolist can account for a fair bit of the data with very few theoretical terms, only a connectionist can account for the rest of the data. Although this requires the invocation of many more theoretical terms, these can be justified if they represent REAL entities in the REAL world (viz., neurons, neural clusters, neural loops, cell assemblies, or what have you). In other words, who needs parsimony if you've got the real thing? (Of course, the problem is in convincing others, particularly other instrumentalists, that you do indeed have the real thing.)
6. To summarize: I offered no justification of the realist approach, not because I believe it to be "superior," but rather because I believe it to be the approach favored by most psychologists, particularly by connectionist cognitive scientists. There is no reason NOT to explore alternative options, such as instrumentalism, but it does not seem to favor the connectionist to do so. On the contrary, connectionists are better served by adopting a realist stance, which gives them a clear justification for the many theoretical entities they typically invoke. Doing this requires connectionists to state just what the nodes and connections in their models represent, and the basic constituents of neural activity seem to be the best candidate currently. New candidates are always welcome, however.
Churchland, P. M. (1990) Cognitive activity in artificial neural networks. In: Thinking: An Invitation to Cognitive Science (Vol. 3), ed. D. N. Osherson & E. E. Smith, MIT Press.
Green, C. D. (1998) Are connectionist models theories of cognition? PSYCOLOQUY 9(4) ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/ psyc.98.9.04.connectionist-explanation.1.green
MacCorquodale, K. & Meehl, P. E. (1948). On a distinction between hypothetical constructs and intervening variables. Psychological Review, 55, 97-105.
McClelland, J. L., Rumelhart, D. E., & Hinton, G. E. (1986) The appeal of parallel distributed processing. In: Parallel Distributed Processing: Explorations in the Microstructure of Cognition (Vol. 1), ed. Rumelhart, D. E. & McClelland, J. L., MIT Press.
Rumelhart, D. E., Hinton, G. E., & McClelland, J. L. (1986) A general framework for parallel distributed processing. In: Parallel Distributed Processing: Explorations in the Microstructure of Cognition (Vol. 1), ed. Rumelhart, D. E. & McClelland, J. L., MIT Press.
Skinner, B.F. (1950). Are theories of learning necessary? Psychological Review, 57, 193-216.
Van Fraassen, B. C. (1980) The Scientific Image. Oxford University Press.
Young, M. E. (1992) Are hypothetical constructs preferred over intervening variables? Commentary on Green on Connectionist-Explanation. PSYCOLOQUY 9(9) ftp://ftp.princeton.edu/pub/harnad/Psycoloquy/1998.volume.9/ psyc.98.9.09.connectionist-explanation.6.young