Genetic Algorithm Enlarges the Capacity of Associative Memory
Akira Imada, Keijiro Araki · Kyushu University Institutional Repository (QIR) (Kyushu University) · 1995
We propose a genetic algorithm for mutually connected neural networks to obtain a higher capacity of associative memory. In Hopfield network as an associative memory system, the memory capacity is at most 15% of the number of neurons. Here we applied our method to the Hopfield network, and obtained the capacity of 33%. We conjectured that this is due to both asymmetry and sparseness of the connection matrix introduced by the genetic algorithm. 1 Introduction In 1982 Hopfield [1] proposed a neural network model as an associative memory system. He used the Hebbian rule [2] to make connection matrix memorize a set of patterns. He estimated experimentally the memory capacity to be 15% of the number of neurons. Later the capacity was verified more analytically by Amari et al. [3] with neurodynamics and by Amit et al. [4] with spin-glass theory. Since Kohonen [5] proposed several types of connection matrices for associative memory, many researchers have successfully improved the capacity in...