How are skills despecialized (or generalized)? One approach would be to replace polynemes with less specific isonomes. But this replacement -- which is an example of abstraction -- must be constrained to prevent absurdities.
"What we call 'generalizing' is not any single process or concept, but a functional term for the huge societies of different methods we use to extend the powers of our skills. No single policy will work for all domains of thought, and each refinement of technique will affect the quality of the generalizations we make."
NB: Generalization is incredibly important for connectionist networks. But connectionist studies of generalization ignore Minsky's insight in the quote above. What would happen to connectionism if it thought of different kinds of generalization, and different techniques for exploring or improving these different kinds? For example, why not consider generalization as a learning effect (e.g., measuring savings) instead of sum of squared error to untrained examples?