| Steven Phillips|
Information Science Division
1-1-4 Umezono, Tsukuba, 305, Japan
| Graeme S. Halford|
Department of Psychology
The University of Queensland
Brisbane, 4072, Australia
|•||Human cognitive behaviour is grouped on the basis of common structure (e.g., from above, it is not the case that one can do the first inference, but not the second).|
|•||Classical architectures capture this grouping of behaviours by positing structure sensitive processes.|
|•||Connectionist architectures, by specifying context- sensitive (structure insensitive processes), distribute behaviour irrespective of structure.|
|•||Therefore, classical (symbol) systems are a better explanation for cognitive architecture, although connectionist architectures may provide suitable implementations of classical ones.|
At issue here is not whether an architecture can ultimately exhibit all the observed stimulus-response behaviours, but how these behaviours are distributed over their available resources (e.g., learning trials). For example, an architecture based on simple associations requires two association steps (e.g., 1: A→B; 2: B→A) to support a bidirectional link between events A and B. By contrast, a relation based architecture only requires one step (e.g., R(A,B)), since bi(omni)directionality is built into relational operators ( Phillips, Halford, & Wilson, 1995). The two architectures, although supporting the same functionality, distribute that functionality differently. The relevant difference is that there are states of associative based architectures for which representations of events are accessible in one direction, but not the other (e.g., after step 1, but before step 2). If one only ever observes bidirectional behaviour then such observations would be support for the relation based architecture, and not the association based architecture, although the former could be implemented by the latter1.
Clearly, then, the root of the systematicity argument over cognitive architecture rests on the degree to which human cognition is systematic. Fodor and Pylyshyn take systematicity to be self-evident. Without recourse to specific data they claim, for example, that one can make inferences of the form P → Q, P ⊢ Q, if and only if one can make inferences of the form Q → P, Q ⊢ P. Subsequently, Hadley ( 1994) characterized systematicity as generalization to novel syntactic position, based on a review of language learning. Researchers have demonstrated networks supporting this definition of systematicity to various degrees ( Christiansen & Chater, 1994; Hadley & Hayward, 1994; Niklasson & van Gelder, 1994; Phillips, 1994). However, others2 question whether the empirical evidence supports this definition either way, given the difficulty of controlling subjects' background knowledge and observing what knowledge they have acquired in the course of an experiment. Furthermore,____________________
Questia, a part of Gale, Cengage Learning. www.questia.com
Publication information: Book title: Proceedings of the Nineteenth Annual Conference of the Cognitive Science Society. Contributors: Michael G. Shafto - Editor, Pat Langley - Editor. Publisher: Lawrence Erlbaum Associates. Place of publication: Mahwah, NJ. Publication year: 1997. Page number: 614.