they can nevertheless be thought of as conceptual components of the
output.
So connectionists sometimes claim that they can accommodate the
systematicity of contentful events in terms of the way in which activation-
patterns can display a distinctive ‘clustering’ appropriate to diVerent con-
cepts. ButWrstly, this just seems to be a brute fact about (some) connec-
tionist networks. There seems to be nothing in connectionism, as such,
which guarantees it, or even makes it likely. And secondly, these patterns
seem causally epiphenomenal. It is not a particular clustering of node-
activations which explains the output and further eVects of a system, but
rather the level of activation across all of the nodes. Thus Fodor has
argued that while such activation vectors may exist, they do not play any
causalrole in the activity of a connectionist network; whereas wedothink
that conceptual components play a causal role in reasoning (Fodor and
Pylyshyn, 1988; Fodor and McLaughlin, 1990).
Some have accused Fodor of committing a sort ofreductionist fallacy
here (Matthews, 1997). The suggestion is that Fodor must be reasoning as
follows: since the real causal work in a connectionist network is done by
the activation levels of the individual nodes, together with the weighting
with which each node transmits its activation to other nodes, apatternof
activation cannot itself be causally eYcacious; rather, it just supervenes on
what does the real ‘pushing and pulling’ within the system, namely the
activities of the individual nodes. Such an argument is fallacious because
the causal activity within such complex systems may operate, and be
correctly describable, at many diVerent levels. Would not the same line of
argument applied to a chemical system, for example, show that chemical
properties must be causally irrelevant, because the real ‘pushing and
pulling’ is done by the underlying sub-atomic mechanisms? (There is some
irony in attributing such a fallacy to Fodor, of course, since he is famous
for defending the reality of, and the causal eYcacy of, the ‘special sciences’
- see his 1974.)
In fact the argument need not presuppose that the real causal work must
be done at a mechanistic level. We (and Fodor) are quite happy to allow
that there is real causality wherever there are events linked together by
causal law, or by dependable nomic (productive) tendency, as we saw in
chapter 7. So we are quite happy to allow that a pattern of activation
might, in principle, be the sort of thing which can causally produce an
eVect, provided that there are someceteris paribuslaws linking the pattern
with its eVects. The crucial point is that in distributed connectionist
systems there is no reason to believe there are any such laws linking
activation-clusters with outputs. In general such clusterings are unstable
and liable to shift when the system is presented with a new range of inputs.
204 Forms of representation