Holograph contraction by oscillatory filtered learning for dynamic sub-pattern matching

The paper presents a scheme for reducing memory space of a holographic associative memory for content-based learning, searching and retrieval of sparse patterns. Multidimensional holographic associative memory developed on the properties of complex valued Riemann space is one of the most promising models of associative memory. It has demonstrated the unique ability to perform dynamically localizable sub-pattern matching, without requiring to learn each individual sub-patterns. The correlation space of the sparse patterns, is also sparse in information, but representationally dense. Therefore, holograph of sparse patterns (such as images) becomes extremely large. In this paper we describe a holographic memory model which can prune a holograph by several fold. The resulting holographic model also simultaneously increases the encoding, searching and decoding speed.