way that only connectionist models can explain. Now, it may be that a
little tampering with the hidden nodes within a network will still leave a
system which can produce imperfect but acceptable output, whereas dam-
age to a serial processor is liable to result in the whole program crashing.
But does this mean that connectionism provides the best model of how
memory functions, malfunctions, and declines? If dissociation in cognitive
performance in other areas is best explained by postulating modules, then
there is clearly an alternative hypothesis. So we think it worth exploring
whether this eVect might not be explained by postulating a memory
system structured out of many sub-modules, which can be independently
damaged.
We do not mind conceding thatsomehuman memory systems may
operate bysuperpositional storage. This is particularly plausible where
memory can consist in recognitional capacities of one sort or another. For
connectionist systems are at their strongest when it comes to pattern-
recognition. (For example, there has been some success in devising connec-
tionist programmes for face-recognition – see O’Tooleet al., 1994.) We
suspect that there is nothing in the folk-psychological account of recog-
nition-memory which requires such memories to be individually represen-
ted. But wedothink that folk psychology is committed to the idea that
cognitiveprocesses(for example, of practical inference or of belief-for-
mation) involve individual (and individually contentful) events – see sec-
tion 2.3 below. Moreover, if there is anything inthe tracking argument
outlined above, then memory will need to be continually updated with
information supplied in a language of thought, and will also itself inform
decision-making processes which require the representational power of
such a language.
In general, too, we would advise caution over allowing that any dis-
tributed connectionist system which can be trained-up to mimic human
performance in some domain of recognition or memory can therefore be
treated as explanatory – that is, as providing acorrectaccount of human
performance in that domain. It is also crucial that the system should share
the same human learning trajectory. So if humans can do one-oVlearning
in the domain in question, for example, then so too must any connection-
ist network which adequately models the human capacity. This constraint
is very commonly ignored or glossed over by connectionist modellers, with
information about training-trials either not given at all, or conWned to an
obscure footnote. But in fact it is vital. Recall from chapter 3 that one of
the main arguments in support of the modularist and nativist research
programme in cognitive science is thespeedof human learning in many
domains, following a common developmental trajectory over highly vari-
able environmental inputs. To the extent that a connectionist system in
Mentalese versus connectionism 199