A proposal for indicating quality of generalization when evaluating ANNs

An expression is proposed to serve as a model for indicating quality of generalization when evaluating ANNs (artificial neural networks). A conjecture made by G.G. Lendaris and G.L. Stanley (Inf. Syst. Sci.; Proc. 2nd Congress, Baltimore Spartan Books, 1965) is repeated which predicts that if an ANN successfully learns a training set, then the smaller the ANN´s performance space, the better will be its generalization. It is argued that the chances of an ANN learning a given task are enhanced if a significant fraction of the possible inputs from the ANN´s input space is in the don´t-care set