Exploring the Symbolic/Subsymbolic Continuum: A Case Study of RAAM
As Smolensky (1988) noted, the term subsymbolic paradigm is intended to suggest symbolic representations that are built out of many smaller constituents: “Entities that are typically represented in the symbolic paradigm by symbols are typically represented in the subsymbolic paradigm by a large number of subsym bols” (p. 3). Smolensky suggested that for the purposes of relating these two paradigms, it is often important to analyze subsymbolic models at a higher level: “to amalgamate, so to speak, the subsymbols into symbols” (p. 3). There is a problem with this type of analysis because a conglomerate of subsymbols does not form a traditional symbol. Classically, symbols have been arbitrary labels, such as strings of letters, which are atomic, discrete, and static. In contrast, a symbol in the subsymbolic paradigm is distributed over a collection of subsym bols, and each subsymbol may be associated with continuous numerical values. In addition, the subsymbolic paradigm is strongly committed to learning at the subsymbolic level. Through learning, an amalgamated symbol gradually emerges in such a way that its form reflects its function, or use, in the training tasks. We see examples of this in the experiments described in section 5.