systems which would provide simulations ofgeneral, as opposed to do-
main-speciWc, learning. But this hope has not been borne out. Most
connectionist networks still require many thousands of training-runs be-
fore achieving the target performance. This contrasts with human learning
which, in many domains at least, can be one-oV– human children will
often need only a single exposure to a new word in order to learn it, for
example. We predict that if connectionism is to achieve real success in
domains which are most plausibly thought of as modular – such as various
aspects of language-learning and language-processing, face-recognition,
categorial-perception, movement-perception, and so on – it will be by
devising networks with a structure which is speciWc to each domain, and
which contain quite a high degree of pre-setting of the weights between
nodes.
3 Developmental rigidity and modularity
In language acquisition and other areas of normal cognition, the one great
and impressive regularity is both the similarity of developmental stages
and the common adult capacity. We maintain that cognitive development
isrigidin the sense that it tends to converge on speciWc and uniform
capacities over a wide range of developmental experiences. The extent of
this convergence is only becoming fully apparent in the light of research
which has revealed the domain-speciWc and modular character of much
cognitive processing.
Nativism and modularity are distinct, in that while nativism is a thesis
about how cognition develops (involving the claim that it is independent
of experiential input, to a signiWcant degree), modularity is a matter of
how cognitive processing is organised. But these two research program-
mes are clearly mutually supportive. For it is not at all plausible to
suggest that the same detailed modular organisation should be replicated
in diVerent individuals simply by the operation of general learning pro-
cesses upon diverse experiential inputs. Modules seem to be special-pur-
pose, dedicated cognitive mechanisms, and one of the major theoretical
arguments in favour of modularity – at least in relation to perceptual
input modules – is that there is adaptive advantage to having cognition
structured in this way. If that is so, then the adaptive advantage will need
to be replicated through genetic transmission of instructions for the
growth of modular systems. In other words, part of the theoretical case
for modularity depends upon nativism being correct at least as far as
some modules – perceptual input modules – are concerned. Further, if
cognitive processing is functionally organised in terms of modules which
are domain-speciWc and also common to the human species, then the best
56 Modularity and nativism