Similarity and pathology in neural nets

Pathological behavior of some recently proposed neural net models is shown to depend on how the similarity is defined. The obvious binary interconnection schemes using weights of (1.0) or (+1, M-1) are shown to give unsatisfactory performance for certain types of stored objects, for instance objects that are subsets of other objects, or for certain types of input distortions. The modified balanced weights model, a modification of a model previously proposed by the author and collaborators (1989) is shown to give the best overall performance and to avoid all the types of pathological behavior considered. The consequences on optical implementations of those networks are discussed