cells, responding diVerentially to particular features, such as the presence
of a colour, or an upright line in a particular region of the visualWeld. In
fact on the basis of what is known so far, it seems quite likely that neural
processing, in general, proceeds by means of the logical operations of
addition and subtraction, in such a way that the representational status of
individual neurons is preserved. So neurons representing straight edges in
diVerent portions of the visualWeld will sum together (minus any input
from neurons representing curvature) to cause a cell toWre which rep-
resents the presence of a straight line, say.
Connectionism is also often supported by pointing out that processing
in the brain is carried out in parallel, not serially as in a conventional
computer. But this is a muddle.Parallelprocessing should not be confused
withdistributedprocessing; and only the latter is inconsistent with some
sort of Mentalese-based account of cognition. In fact, anyone who believes
in modularity should accept parallel processing. If the visual system sub-
divides into a number of distinct modules, for example, each of which
processes its input independently of the others before passing on an output
for integration, then the visual system will process inputs in parallel. But it
is quite another matter to claim that the processing which takes place
withina given module or sub-module contains no states which represent
independently of the others.
So it is important to distinguish betweenparallelprocessing anddis-
tributedprocessing. If there is a contrast to be drawn between symbolic and
connectionist systems, that contrast shouldnotbe seen as one between
linear and parallel forms of processing. The fact that the brain is known to
process contents in parallel, in many domains, provides no particular
support for the connectionist approach. It is perfectly possible for sym-
bolic systems to process contents in parallel, by devolving processing to a
variety of modules or sub-modules, each of which operates independently
of the others. As we saw in chapters 3 and 4, the case for a modular
structure to cognition (including central cognition) is a powerful one. So
we are happy to allow that processing of perceptual inputs will be conduc-
ted in parallel; and that central processes of thought and reasoning, too,
are often conducted in parallel, devolved to a variety of conceptual mod-
ules. All this is quite consistent with those processes involving computa-
tions upon symbols, or symbol-like structures. Matters are otherwise,
however, if connectionism is understood as proposing that processing is
distributed across a network, computing patterns of activation across the
nodes, in such a way that none of the computations in question can be
described as transformations of symbols.
Connectionism is also sometimes supported by pointing out that human
memory systems degrade gracefully, with gradual loss of function, in a
198 Forms of representation