A hierarchical associative memory consisting of multi-layer associative modules

The author proposes a hierarchical associative model consisting of multilayer neural associative memory modules which do not have any intra-layer connection and which serve as memories for micro-level retrievals. The constituent memory modules are connected to each other by having their layer in common. The hierarchical architecture decreases the total number of connections as compared with the previously proposed models. The hierarchical model has been applied to a character-word recognition problem and the interaction between the constituent modules has been studied. The success rate of the recognition has been found to be controlled by the model´s parameters.