The ideal way, of course, to solve the stability-sensitivity problem would be to store all
previously learned patterns “out of reach” of the disruptive influence of new patterns. Then,
when new input was presented to the system, all of the previously learned patterns would be
taken out of storage, so to speak, and would be mixed with the new patterns. The system
would then learn the mixed set of old and new patterns. After the augmented set of patterns
had been learned by the network, they would all be put in storage, awaiting the next time new
information was presented to the network. There would be no forgetting, catastrophic or
otherwise, in this ideal world and there would be no deleterious effect on the network’s
ability to generalize, categorize or discriminate. The system could learn the new patterns and
the old patterns would be unaffected by the acquisition of this new information