Content-based information retrieval using an embedded neural associative memory

Abstract

In this paper a novel approach for the storage and access of an index used in Internet search engines (Information Retrieval) is presented. The index provides a mapping from search terms to documents. The Binary Neural Associative Memory (BiNAM) stores an index by associating document signatures and document locations in a distributed and content addressable way. The system presented here has a high memory efficiency of more than 90%. The trade-off between memory consumption and precision of the query-results is examined. A scalable system architecture is presented. The architecture exploits the parallel structure of the BiNAM. The association time is estimated to be orders of magnitude faster than a software solution. The system is realized as a modular PCI architecture. The maximum capacity of the first version is 768 MByte memory which allows to implement a BiNAM of 80 K neurons with 80 K inputs each. In such a system approximately 64 million associations can be scored and accessed within 330 ns per association.