A content-addressing software method for the elimination of neural networks

A software method is generally suitable for the implementation of neural network structures is described. In the particular text-processing examples discussed, the nodes correspond to local features that are made content-addressable by a special use of hash coding. Higher-level associations are encoded by logical pointers between features. Such a network structure can store large amounts of discrete information, e.g. text, and accomplish a very effective proximity search activated by erroneous or incomplete cues. The method can be implemented using conventional programming techniques, and it has been successfully applied to isolated-word speech recognition and sentence reconstruction.<>