ABSTRACT
Connectionist systems have shown themselves adept at a certain kind of problem solving. They are adept at delicately trading off multiple soft constraints so as to behave as if they encoded symbolic knowledge of rules and categories (examples to follow). In so doing, they are getting by with emergent and example-bound correlates to classical rules and data-structures. I shall argue that this emergentist style of representation, despite some clear benefits of interpolation, generalization and content-addressability, is unable to provide a distinctive kind of cognitive flexibility exhibited by human subjects. (I leave the question of the animal kingdom to the experts: probably some higher animals exhibit similar flexibility.) In short, I shall argue for the poverty of the purely emergent.
