And even if an activation-cluster remains suYciently stable to support
some generalisation linking patterns of activation with outputs, this is a
fact about how a particular network is functioning, rather than a regularity
of connectionist processing which might have law-like status. For the
patterns of clustering are a product of the inputs in the training set and the
method of adjusting weights and biases in response to information about
the output, and what the patterns are ishighly sensitiveto the exact inputs
and method of adjustment. So here too one needs to consider the details of
the training regime, rather than just the performance of a single specimen
of a trained-up network.
2.5 Connectionism, holism and the frame problemHorgan and Tienson (1996) draw an important distinction between two
diVerent, and partially independent, commitments of classical (that is,
symbolic;non-connectionist) approaches to cognitive science. On the one
hand, classicists are committed to the idea that propositional attitudes
havesyntactic structure– that is, to the idea that there is a language of
thought (LoT). And on the other hand, they are committed to the claim
that cognitive processing isrule-governed, operating in accordance with
strict (albeit probabilistic) algorithms. It seems very likely that the second
of these ideas entails theWrst – that is, that algorithmic processing must
operate on structured states. But theWrst does not entail the second – while
cognitive states are syntactically structured, it could be that transitions
amongst such states are not rule-governed but rather, in some sense,
‘chaotic’. Horgan and Tienson argue at length for the claim that proposi-
tional attitudes are syntactically structured states. As we have seen, they
argue that we cannot make sense of complex intelligent cognition without
supposing that mental states are systematically built out of conceptual
components. But they also argue at length against the classical assumption
ofalgorithmicity. Rather, they think that the correct way to model and
explain mental-state transitions will be by using some or other form of
dynamical systems theory.
We are happy to endorse this distinction. And we, too, want to insist
that propositional attitudes are compositionally structured states. This in
itself is enough to ensure the reality of the states postulated by folk
psychology, since the folk are certainly not committed to anything con-
cerning the algorithmic nature of the processes which transform and
generate such states. Moreover, it does seem to us to be an open – and
entirely empirical – question whether cognitive processing is algorithmic,
or would be better modelled by a diVerent branch of mathematical theory,
such as dynamical systems theory. But we do think that theargument
Mentalese versus connectionism 205