Semantic Neuron Networks Based Associative Memory Model

In our previous work, Hopfield weight matrices were used as the primary inputs to learn associations among concepts or objects in different learning tasks. One of the challenging problems, however, was how to endow the final results of the system with appropriate semantics so as to enable easy retrieval of associated memory items in the course of continuously learning and experiencing like we humans do. This paper proposes a semantic neuron network based associative memory model in which each matrix representing a resulted Hopfield network in a learning task is endowed with a semantic by the generated semantic net. Chunking mechanisms performed on matrices are proposed to be the means through which merging and decomposition of correlated matrices is done. This study is aimed at developing an associative memory model which can self-organize and self-evolve in its lifetime. This process is quite similar to the human brain activity which has been shown to use associations when forming complex memories.