There is a tendency for advocates of one of these two computational cognitive paradigms to assert that it is better than the other, usually on the grounds that some cognitive behavior has been implemented more or less successfully in the one but not (or not yet) in the other. There is also a growing tendency for some researchers to seek a ``meeting of the minds'', usually on the grounds that the cognitive behaviors that have been implemented more or less successfully using the one paradigm are precisely the ones that have not (yet) been successfully implemented using the other. In particular, cognitive processes that are easy to implement symbolically (e.g., problem solving, reasoning, game playing, certain aspects of linguistic competence) tend to be ones that are relatively difficult for humans or that have to be explicitly taught, while those that have proven difficult to implement symbolically (e.g., certain aspects of visual perception and learning) tend to be those that ``come naturally'' to humans. This paradox of (symbolic) artificial intelligence has its counterpart in the debate over connectionism. The processes that have proven difficult to implement symbolically appear to be susceptible to connectionist techniques. This complementarity may prove to be a major advance in our understanding of cognition.
A second way of merging the two approaches is to view connectionism as a lower level of cognitive processing, i.e., as a way of implementing symbolic processes. Thus, e.g., logical reasoning, which is well-suited for symbolic computing, might be implemented in a connectionist system in such a way that certain of its ``connectivity patterns'' reliably represent precisely the things that would be explicitly represented symbolically. Some proponents of the PSSH would put this slightly differently. They would say that at some stage in the sequence of levels that describes a computer (beginning with the ``device level''--the description in electronic terms), there must be a level that implements a symbol system. (Cf. Newell 1981: 76, Pylyshyn 1985: 96, Clark 1990.) Several connectionist implementations of ``symbolic'' algorithms have been investigated, but as yet there is no general theory of how this might be accomplished.