Symbols and subsymbols for representing knowledge: a catalogue raisonne

Traditional artificial intelligence studies generally approach the problem of representing knowledge following the so-called knowledge representation hypothesis, as formulated by Brian Smith. More recently the development of the connectionist paradigm has questioned the symbolic approach to the study of the mind bringing about a more articulated view of the problem. This article singles out five possible approaches to the problem of knowledge representation in cognitive science: compositional symbolic approaches, local non-compositional approaches, distributed non compositional approaches, cognitive subsymbolic approaches and "neural" subsymbolic approaches. In particular, in the subsymbolic cognitive approach the elements that make up the representation system are not symbols with an ascribed meaning nor do they correspond to anatomic entities at a neurological level; rather they are to be considered as "theoretical constructs" of a theory of the cognitive level which permit the deduction (in the sense of "computation") of cognitive behaviours which cannot be otherwise modelled. We consider the development of models of this kind to be essential to a computational approach to the problem of reference without hypothesizing "magical qualities" of the mind (in the sense of assuming a necessary connection between mental symbols and their referents), while remaining within a functionalist vision, in the wider sense, which does not make reference to the specific physical properties of the neural hardware.