Connectionist models of language: Learning from successes and failures that are not failures

Seidenberg, M.

Are PDP models relevant? Many of the basic principles underlying the approach have been assimilated into contemporary thinking in cognitive science and neuroscience, but the modeling methodology has not. This is too bad. Interest in PDP models took hits from analyses by McCloskey (1993), Pinker (1991) Marcus (2001) and others. Upon further review (as they say in American football), it is astonishing how little of what was claimed about the limits of PDP models turned out to be true (catastrophic interference? The past tense? Algebraic rules? No.). The modeling aspect of PDP was open to a deeper concern, however: taking credit for the models' successes and assigning deviations from reality to implementational details. What good is a model if it can never be wrong? I'll argue that in this respect PDP models are like every other approach in cognitive science ranging from Chomskyan linguistics to cognitive Bayesian models. Finally, I'll show that the failure of PDP models to penetrate neuroimaging research has led to an unfortunate revival of box-and-arrow theorizing -- presenting an opportunity for important new modeling research.