Rather, modules areimplicitlydifferentiated, by what formats of cognitive structure theyaccess and derive. Such a view
of modularity has been espoused by Townsend and Bever (1982), Jackendoff (1987),^110 Arbib (1987), Levelt (1989),
Prather et al. (1991), Bever (1992), and Coltheart (1999), among others. As we will see, it provides a natural
explanation of many of the phenomena with which Fodor is concerned.^111
Each module is strictlydomain-specific in Fodor's sense: integrativeand inferential processors deal with only onelevel
of structure each; interface processors deal with two (we might therefore want to call them“bi-domain-specific”).
Similarly, each module is informationally encapsulated: the only kind of information that can influence it is its
designated input level. Through the chaining of integrative and interface processors—and the possibilities for
constrained feedback among them—we achieve the overwhelmingly complex mapping between acoustic information
and meaning. Furthermore, if each processor is mandatory and fast, then the chain will be mandatory and (almost as)
fast. That is, the effect of Fodor's faculty-sized module is created by the chaining of a series of structure-specific
modules.^112
Much of the dissent concerning modularity (e.g. Crain and Steedman 1985;
220 ARCHITECTURAL FOUNDATIONS
(^110) In Jackendoff(1987; 1997a; 2000) I used theterm“representational modularity.”The new terminology is adopted hereas part of theeffort to expunge“representation”
fro mthe theoretical vocabulary, for reasons discussed in Ch. 2.
(^111) A somewhatdifferentscaleof modularityis invokedin theworkofCosmides and Tooby(e.g. Cosmides 1989; Toobyand Cosmides 1992) , who sometimes speak ofsuch
modules as“cheater detection,”“friendship,”and“kin-oriented motivation”. My inclination is to regard these as too small a domain for modularity: each of them is
embedded in a richsystemof cognitivestructures and inferentialprocesses concerningvaluesand socialinteraction. Itis thislarger systemthat(tome at least)appears tobe
on a more appropriate scale for modular processes in the present sense (Jackendoff 1992) .8; ch. 4; 1994: ch. 15). Within such a system, cheater detection and so forth
mightbe regarded as principlesof socialinferencewhose acquisitionis highlyfavoredby thegenome, more or less akin to the“attractors”of UniversalGrammar proposed
in section6.10. It appears to me that such a perspectivedoes no damage to the genuineresults of the Cosmides-Toobyprogram, but has the effectof making the rhetoric
more compatible with that for language.
(^112) We returntothe“slow”centralprocessespositedbyFodor,whichapplytoconceptualstructure,inChs. 9 and 10). To theextentthatFodoriscorrectaboutbelieffixation
being relativelyslow, isotropic, and Quinean, these properties are now confined to inferential processors which compare newly constructed conceptual structures to those
present in long-term memory, and which construct further inferences and judgments on the basis of this comparison. If, as I believe, there is a furthersyste mof central
cognitive structure concerned with spatial understanding (Ch. 11) , this too may well support slow and isotropic inferential processes: consider spatial problems like
optimally packing a carton, designing a better mousetrap, and deciding the best route to drive to the other side of town. Such a division between propositional/conceptual
structures and spatial structures is a central aspect of Stephen Kosslyn's (1980; 1996) research program. The upshot is that central processes may well be modular too,
though of a different sort than the input and output processes.