What have neural networks achieved?

[Gary Cottrell]

(Hi Jay)

I like the Seidenberg & McClelland account of reading, as
well as PMSP. However, the PMSP result (nonword reading)
really relies on a very well engineered representation of
the outputs. That is, it is very difficult to get a poor
pronunciation with that representation, which already encodes
many of the rules of English pronunciation. Thus, it seems
like one of Lachter and Bever's TRICS (The Representation It
Crucially Supposes). This does not contradict the fact that
a single mechanism account has been demonstrated. But one of
Jay's "loose ends" is: How would a network develop such a
representation in the first place? Which I think is a crucial
question.

g.