Connectionism, Analogicity and Mental Content
11
In summary, the standard PDP networks discussed in the literature are analog
computational devices. And if PDP networks are analog devices, then connectionism, the theory
one gets when one applies the PDP computational framework to cognition, is an analog
conception of mind.
5. The Cost of Connectionism: A Resemblance Theory of Mental Content
According to H&T, what sets connectionism apart from classicism is its eschewal of the discrete
mathematics of digital computation in favour of the fundamentally continuous mathematics of
dynamical systems theory; cognitive processes are not governed by exceptionless,
“representation-level” rules, they are the work of defeasible cognitive tendencies subserved by
the non-linear dynamics of the brain’s neural networks. (1996, Chps.2&4). To this extent, their
dynamical characterisation is quite consistent with the analog reading of connectionism
developed in the previous section. But dynamical systems theory, on its own, is not a
computational framework. There is no fundamental difference, from a purely dynamical
perspective, between NETtalk and the convection currents generated in a pot of boiling water. In
this sense, dynamical systems theory dissolves the distinction between intelligent and
unintelligent behaviour, and hence is quite incapable, without supplementation, of explaining
cognition. In order for dynamical engines to be capable of driving intelligent behaviour they
must do some computational work: they must learn to behave as if they were semantic engines.
H&T recognise this fact. They recognise that as classicists have a robust story to tell about
how semantic coherence is achieved in cognition (or “content-appropriateness” to use their
term), connectionists must do the same. The story they subsequently tell (see especially 1996,
Chps.6&9) focuses on the role of “content-appropriate cognitive forces”, subserved by spreading
activation across PDP networks. “Certain kinds of content-relevant interaction are automatic for
systems that have states that emit content-relevant cognitive forces”, H&T claim, which is a key
difference from classical systems “in which the operative PRL [ie, programmable
representational level] rules must determine all such outcomes” (1996, p.99). But exactly how are
such content-appropriate forces realised in PDP networks? The answer H&T provide is that in
training up a network, the representations that comprise cognitive states are placed strategically,
rather than arbitrarily, on an activation landscape, such that “their relative-position relations
systematically reflect key semantic properties and relations of the relevant individual cognitive
states themselves” (1996, p.156).
This is where H&T leave it. But one is entitled to ask for more. In particular, one is
entitled to ask how an activation landscape could be so arranged that its various regions reflect
the semantic relations to which H&T refer. The only answer that would seem to be available
here, the only answer that entitles connectionists to dispense with representation-level rules, is
one that invokes a structural isomorphism between this activation landscape and its
representational domain. It’s this isomorphism that renders the shape of the activation landscape
semantically significant, and hence endows the points in this landscape with the content-
appropriate causal powers. Dynamical engines, left to their own resources, might be incapable of
engaging in semantically coherent behaviour; but once these engines are coupled to an analog of
some region of the world, they are transformed into powerful computational devices. H&T’s
dynamical characterisation of connectionism is thus based, tacitly if not explicitly, on analog
computation.
The implications for connectionist cognitive science are many. Most striking of all, I
think, is that connectionism brings with it a new focus on the representational vehicles implicated
in cognition, in addition to the conventional focus on the computational processes in which they
are involved. This is because the semantic coherence of cognition, according to connectionism, is
actually embodied in these vehicles, rather than in a set of independently defined computational