This paper deals with the question: what are the implications of connectionism for theories of computation? Three possible answers are examined. 1. A theory of implementation: connectionist representations are not semantically structured. Connectionism can be deemed a theory for implementing classic symbolic computation. 2. Representational connectionist computation: connectionist networks compute by exploiting relations of structural resemblance between their connection weights and their target domains. An adequate representational theory of computation will also explain connectionist computation. 3. Non-representational connectionist computation: connectionism need not be committed to internal representations. An adequate non-representational theory of computation could account for connectionist computation. To some, connectionism seems like a promising alternative to the classical computational theory of mind. This debate will not be pursued directly in this paper. Rather, by critiquing the answers above, it will be examined whether connectionist networks (or neural networks) compute and if so, how they compute1 . This will also be explored by reviewing some examples of neural networks including some recent patents (e.g., colour categorisation network, NETtalk, pattern recognition networks etc.). Moreover, I argue that an important question that should be asked is whether connectionist computation qualifies as digital or analogue computation.
Connectionism, digital computation, analogue computation, cognitive science, representations, implementation, neural networks
School of history and philosophy, The University of New South Wales, Sydney, NSW, Australia.