Content-sensitive inference, modularity and the assumption of formal processing

Abstract

Performance on the Wason selection task varies with content. This has been taken to demonstrate that there are different cognitive modules for dealing with different conceptual domains. This implication is only legitimate if our underlying cognitive architecture is formal. A non-formal system can explain content-sensitive inference without appeal to independent inferential modules.

Acknowledgments

Special thanks to William Bechtel, Jay Garfield, Mark Colyvan, Tim Bayne, Jane French and two anonymous referees for comments on earlier versions of this paper.

Notes

Fodor (see, e.g., Fodor & Pylyshyn, 1988) has raised what he believes are strong independent grounds for rejecting connectionism as an account of human cognitive processing. Similarly, though far less persuasively, Carruthers (2003a, 2003b) suggested that classical computationalism is the only serious game in town. Cosmides and Tooby, however, do not offer any reason for their (implicit) commitment to symbolism.

My discussion of the evidence for modularity closely follows Carruthers (2003b). Fiddick, Cosmides, & Tooby (2000) also mention studies in economic theory, which appears to lack serious empirical support.

Strictly, the first experiment by Wason using the ‘if p then q’ locution to demonstrate substantial reasoning errors was performed in 1964. This experiment did not, however, employ a selection task, but required subjects to answer a simple modal question.

Though most connectionists avoid modularity there are a variety of modular connectionist networks in use. The most commonly employed are manually designed expert systems (see Husken, Igel, & Toussaint, 2002; Jordan & Jacobs, 1995; Sharkey, 1996). Miikkulainen and Dyer (1989, 1991) have also done extensive work on modular connectionist models of language processing.

In a local network, the processing primitives (nodes of the network) and the representational primitives coincide. (The coincidence of representational and processing primitive defines what it is to be symbolic). In a distributed network, the processing and representational primitives do not coincide, as content is attributed to patterns of nodes. Nevertheless, as each node is only associated with a single content, processing proceeds in an essentially symbolic manner. For more detail see Clark (1993).

The discussion of context-sensitivity that follows is based on the recent work of Boden and Niklasson (2000).