Synthesis of associative memories using complex-valued neural networks

Recently a complex-valued neural network has been proposed and applied for implementing associative memories. Its potentials have been investigated mainly from the point of view of storage capacity and recalling ability. But, most of them are direct extensions of well-known autocorrelation type associative memory on real-valued networks. This paper discusses a method for synthesizing associative memories in complex-valued neural networks. The method guarantees that all the desired memories are successfully stored and correctly recalled in the sense that they all become asymptotically stable equilibrium points. In order to enhance the capability of implementing associative memories, we propose a new network architecture of complex-valued neural networks in which dynamic neurons and static neurons are fully connected. We also propose a learning method for synthesis of associative memories by using static networks for the purpose of efficient learning. The proposed learning algorithm makes it possible to realize specified asymptotically stable equilibria in the complex-valued neural networks