This is true, as far as it goes, for any pre-given computation and
given infinite resources (much as Turing machines emulating other
Turing machines can require infinite tapes). In more realistic
senses, however, its typically faster to use a Turing machine to model
a function you know ... a function you don't is often doable with a
learning NNet /of sufficent size and complexity/ faster than trying to
figure out the formulea and coding it up.
That still doesn't mean that there's `computation' going on inside the
NNet, just as there isn't really `planning' going on inside a
spreading activation agent network. Certain things threshold at
certain times and certain actions/behaviours result.
ObMemetics: This may point to an unconscious point of slop in the
terms we use without thinking about it. We use `thinking' a lot, and
seem to imply some active reorganization of memes in the process, but
that's not the only sort of thing that could mechanically be going on
down in there. Again, memetics makes a very handy abstraction as long
as you recognize it as such and /use/ it as such.