Christopher D. Green (1998) Connectionist Nets are Only Good Models if we Know What They Model. Psycoloquy: 9(23) Connectionist Explanation (20)

Volume: 9 (next, prev) Issue: 23 (next, prev) Article: 20 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 9(23): Connectionist Nets are Only Good Models if we Know What They Model

Reply to Lee et al. on Connectionist-Explanation

Christopher D. Green
Department of Psychology
York University
Toronto, Ontario M3J 1P3


Lee, Van Heuveln, Morrison, & Dietrich (1998) suggest, incorrectly, that I argued (Green 1998a) that connectionist networks will not be scientific models unless and until they capture every aspect of neural activity. What I argued was that unless and until connectionists come to terms with the idea that connectionist networks must model SOMETHING (and neural activity currently seems to be the best candidate, but it need not be the only one) they are not models of anything at all, and therefore may have little role to play in cognitive science.


artificial intelligence, cognition, computer modelling, connectionism, epistemology, explanation, methodology, neural nets, philosophy of science, theory.
1. Lee, Van Heuveln, Morrison, & Dietrich (1998) argue that connectionism "properly practiced" can lead to good scientific models of cognition. I think they might be right, but I do not think they have come completely to terms with what "proper practice" might mean for connectionists. Early on in their commentary (para. 3), they seem to suggest that I (Green 1998a)"equate models and theories," and that this is not a "useful way to talk" in computational cognitive science. I agree that it is a simplification, but I thought it best not to enter into the morass known as the "semantic approach" to theories (viz., the one that emphasizes the importance of models over theories) because there are as many views of the relationship between models and theories expressed there as there are scientists and philosophers of science who have considered the question (see, e.g., Suppes 1960; Van Fraassen 1987; Giere 1988; and especially Downes 1992, who comes to the "deflationary" conclusion that models are important in science, but for no single identifiable reason in particular). I find no reason to believe that Lee et al. have solved this decades-old problem either. Their claim that "the theory specifies the similarity relations that obtain (assuming they mean 'are supposed' to obtain) between the model and the phenomenon to be explained" is a well-worn one, but not one that has historically been able to hold much water. What is more to the point, they never tell us what, exactly, they think the details of connectionist theory might be such that it is able to tell us what similarity relations obtain between connectionist networks and cognition. Indeed, I don't see how one COULD specify such a theory unless one knew to what aspects of the phenomenon the theoretical terms of the model refer (which is just a reiteration of my original argument).

2. The core of Lee et al.'s argument seems to begin with their assertion that "a model does not have to be isomorphic to the phenomenon concerned in all respects. Instead, all that it is required is to share relational similarities with the phenomenon in certain RELEVANT respects" (para. 5). This is, of course, a truism with which no one -- certainly not I -- could disagree. They then go on to use it as a foundation for their further argument that "from the fact there are many differences between biological neural networks and connectionist networks it does not follow that connectionist models are not scientific models, nor does it follow that the theories in which they participate are not scientific theories" (para. 5). All this is true, but beside the point. My claim was not that connectionist nets aren't "scientific" (I'm not sure I even know what they mean by this) because they aren't "isomorphic" with neural activity. My claim was that unless and until connectionists are willing to accept specifically that their networks are literal models of neural activity (or of any other concrete domain of phenomena) -- e.g., that nodes are neurons, though as I've repeated many times, it needn't be that crude a mapping -- we will not know what they are modeling at all.

3. Lee et al. (para. 6) go on to discuss post hoc statistical analyses. I have already discussed these in a reply to Medler & Dawson (1998; Green 1998b). Lee et al.'s claim that such analyses "are not, in general, ad hoc" is an interesting one. In general, connectionists seem to conduct descriptive rather than confirmatory factor and cluster analyses on their networks, and to that degree they are ad hoc. Indeed, it was the legendary "ad hoc-ness" of such analyses that contributed to the poor reputation such statistical methods now have among many research psychologists, and it was why they have generally opted, in recent years, for LISREL and structural equation modeling techniques instead.

4. Lee et al. also criticize my argument that connectionist networks have too many degrees of freedom by attempting to draw an analogy between connectionist networks and model airplanes. Does the fact that a model airplane in a wind tunnel has many degrees of freedom, they ask, "mean that the model airplane is of no use in understanding the behavior of an actual air plane" (para. 7)? I reject the analogy outright. We have good reason to believe that a model airplane operates on the same aerodynamic principles as does a real airplane: they both have wings, tails, flaps, etc. The whole point of my target article was that we currently have no reason to believe that such a mapping holds between connectionist nets and cognition at all because we don't know what the nodes and connections correspond to in the cognitive domain.

5. Last, Lee et al. claim that Clark (1993) has shown us how to correctly interpret connectionist research; but they do not respond to the criticisms I have made of Clark's position in previous replies (Green 1998b, 1998c). Until they do, I continue to be dubious that Clark has shown us anything of the kind.


Clark, Andy (1993) Associative Engines. Cambridge: MIT Press.

Downes, S. (1992) The importance of models in theorising. Proceedings of the Philosophy of Science Association (Vol. 1, pp. 142-153).

Giere, R. (1988) Explaining Science: A Cognitive Approach. Chicago: University of Chicago Press.

Green, C.D. (1998a) Are connectionist models theories of cognition? PSYCOLOQUY 9(4)

Green, C.D. (1998b) Statistical analyses do not solve connectionism's problem: Reply to Medler & Dawson on Connectionist-Explanation. PSYCOLOQUY 9(15)

Green, C.D. (1998c) Problems with the implementation argument: Reply to O'Brien on Connectionist-Explanation PSYCOLOQUY 9(8)

Lee, C., van Heuveln, B. Morrison, C.T., & Dietrich, E. (1998) Why connectionist nets are good models: Commentary on Green on connectionist-explanation PSYCOLOQUY 9(17)

Medler, D. A. & Dawson, M. R. W. (1998) Connectionism and cognitive theories: Commentary on Green on connectionist- explanation. PSYCOLOQUY 9(11) psyc.98.9.11.connectionist-explanation.8.medler

Suppes, P. (1960) A comparison of the meaning and use of models in mathematics and the empirical sciences. Synthese, 12, 287-301.

Van Fraassen, B. C. (1987) The semantic approach to scientific theories. In: The Process of Science: Contemporary Philosophical Approaches to Understanding Scientific Practice. ed. by N.J. Nersessian. Dordrecht, M. Nijhoff.

Volume: 9 (next, prev) Issue: 23 (next, prev) Article: 20 (next prev first) Alternate versions: ASCII Summary