Connectionism, Analogicity and Mental Content
representational vehicles participating in a computational process must be semantically coherent,
in the sense that they bear comprehensible (ie, non-arbitrary) semantic relations to one another.
With these emendations in place, we arrive at a very general conception of computation
as a procedure in which representational vehicles are processed in a semantically coherent fashion.1
When this conception is applied to cognition, we arrive, in turn, at the generic version of the
computational theory of mind: cognitive processes are semantically coherent operations over
neurally implemented representational vehicles. This is a very good idea; it is the idea,
moreover, which forms the foundation of cognitive science. But it is an idea that needs to be
fleshed out; we need to know how computational processes, so understood, can be physically
realised. This is where Turing does become important. For Turing showed us a way to bring this
idea to life.
3. Turing’s Way: Digital Computation and Classical Cognitive Science
A computational device is one that mechanistically processes representational vehicles in a
semantically coherent fashion. But how can such a device be constructed? How, to put this
another way, can the semantic properties of physically realised representational states shape the
mechanistic processes by which they are processed? The short answer here is that they can’t -
semantic engines are impossible. The trick to mechanising computation, therefore, is to align the
physical properties of a physical device with semantic ones, such while its causal behaviour is
entirely determined by the former, it nonetheless respects the latter. The trick to mechanising
computation, in other words, is to make a physical engine behave as if it were a semantic one.
Turing saw a way of doing this. What is needed, he reasoned, is a systematic means of
physically encoding information, and a mechanism for manipulating this representational
medium in a semantically coherent fashion. The first requirement, he thought, could be satisfied
by written symbols, and the second by a mechanism that recognises and transforms these
symbols purely on the basis of their material properties, but in accordance with rules that ensure
that these mechanical symbol manipulations are semantically coherent. Turing’s solution was
the specification of an abstract machine - the conceptual basis of the digital computer - replete
with a tape on which symbols are written, and a read/write head so configured that it behaves
as if it were following a set of primitive computational instructions.
Physically implementing Turing’s abstract machine is no easy feat, however. One must
first find a means of physically realising a symbolic representational medium. This is achieved
by systematically partitioning some continuously variable physical property, and then providing
these partitions and their concatenations with a semantic interpretation (whether this
interpretation concerns numbers, propositions, or whatever). The partitioning of a continuously
variable physical property generates syntactic structure, which under interpretation becomes a
symbolic medium. In conventional digital computers this role is performed by the electrical
voltages2 between pairs of wires (eg, strips of aluminium printed on silicon chips), which receive
a binary partitioning into “high” and “low” states. Thus the symbols in such computers are
1 This more general conception of computation can be found, if one looks carefully enough, in a number of places in
the philosophical literature. See, eg, Cummins and Schwarz, 1991, p.64; Dietrich, 1989; Fodor, 1975, p.27; and Von
Eckardt, 1993, pp.97-116.
2 Strictly speaking, electrical voltage is a continuously variable physical magnitude, where a physical magnitude is a
function that systematically assigns numbers to a physical property or properties of a physical system. But certain
computers use continuously variable physical properties to physically implement their representational media that are
not easily characterisable in these terms. That is, the variability of a particular physical property is not always
comfortably captured in a straightforwardly numerical fashion. To achieve the necessary generality, therefore, I will
talk in terms of the physical properties that realise the representational media of computational systems, rather than in
terms of physical magnitudes that numerically characterise these properties.