Derthick, M. and Plaut, D. C. (1986). Is distributed connectionism compatible with the Physical Symbol System Hypothesis? Proceedings of the 8th Annual Conference of the Cognitive Science Society (pp. 639-644). Hillsdale, NJ: Lawrence Erlbaum Associates.

Download: pdf (6 pages, 21 Kb)

Abstract: All existing intelligent systems share a similar biological and evolutionary heritage. Based on the conviction that cognition is computation, artificial intelligence researchers are investigating computational models as a means of discovering properties shared by all intelligent systems.
      One property that has been proposed as central to intelligence is the ability to construct and manipulate symbol structures. If intelligence may be described completely in terms of symbol processing, then cognitive science need not be concerned with the particular physical implementation details of either artificial or biological examples; neuroscience would no longer be part of cognitive science. On the other hand, if important aspects of intelligence evade symbolic explanation, it may prove necessary to consider phenomena below the symbol level. The connectionist approach to artificial intelligence is founded on the conviction that the structure of the brain critically constrains the nature of the computations it performs. However, if the symbolic position is correct and neural networks only implement symbol systems, then connectionism contributes little to cognitive science.
      The notion of intelligence as symbol processing was made explicit by Newell and Simon with the Physical Symbol System Hypothesis (PSSH) (Newell, 1976; 1980) and the Knowledge Level Hypothesis (KLH) (Newell, 1982). Taken together, these hypotheses have significant implications for the nature of any system capable of general intelligence. We examine a number of connectionist systems in light of the hypotheses and distinguish three kinds: (1) rule-based systems, which are symbol systems; (2) rule-following systems, which are symbol systems only under a weakened version of the PSSH; and (3) systems which are not rule-following, and thus are not symbol systems even in a weak sense.
      According to the PSSH, non-symbolic connectionist systems must be incapable of general intelligence. There are strong arguments both for and against this conclusion. On the one hand, such connectionist systems may provide more parsimonious accounts of certain cognitive phenomena than do symbolic approaches. On the other hand, these connectionist systems have significant limitations, relating to universality, not shared by symbol systems. We conclude that a comprehensive theory of intelligence may require a hybrid model that combines the strengths of both approaches.

Copyright Notice: The documents distributed here have been provided as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.