The paper looks at how the hidden-vector cluster analyses associated with Elman and others seemed to provide a potentially important link between the symbolically-oriented level of analysis and the connectionist level of analysis - a link that might one day help to explain how higher mental processes are grounded in neural architectures. The paper goes on to reconsider the implications of these analyses in light of some recent work by Finch and Chater which shows that linguistically meaningful categories (of the type derived from hidden-vector analyses) are directly evidenced in the N-gram statistics of natural language. The implication of this work seems to be that hidden-vector analyses do not primarily address the link between the symbolic and connectionist levels of explanation but rather tell us something about the statistical properties of the training environments used. The consequences of this result for cognitive science are lightly sketched in.
Download compressed postscript file