What do connectionist simulations tell us?

Abstract

In his review, Rispoli's main concern is that Elman et al.'s book will aggravate the degree of polarization in developmental psycholinguistics. I cannot really comment on this worry, as developmental psycholinguistics is not my field. Instead, I will discuss some questions more related to my background – the role of computational modelling in Elman et al.'s approach.

Elman et al.'s ambitious goal is to propose theories of cognitive development that are grounded in our knowledge of biology. This is of course what the great Jean Piaget tried to achieve during his lifetime – unsuccessfully, as we know. Elman et al.'s advantage over Piaget is that they have a set of computational tools, connectionism, which both allows them to specify theories precisely and to study complex behaviours (such as epigenesis, where innate and environmental factors interact to create new levels of complexity) that are just beyond the (unaided) human mind. Even though I will highlight some of the weaknesses of their approach below, I should emphasize that reading their book was an exciting and enjoyable experience.

As noted by Rispoli, there are important problems with the simulations reported by Elman et al. Rispoli focuses on simulations of past tense acquisition and syntax acquisition, but the problems are by no means limited to these areas. I will briefly consider two recent developments in neural net research, one taken from the field of language acquisition, and one from elsewhere, which underscore some of the difficulties of the simulations discussed in the book.