Analysis And Research Of Extension Of Hopfield Neural Network

Since J.J.Hopfield's seminal papers in 1982[1], discrete Hopfield neural network[1][2] have been put forwards,which have been studied extensively[3][10], owing to its simplicity in the network architecture and its fast convergence property. Recently, E.M.Casermeiro shows a new type Hopfield multivalued recurrent neural model MREM[11], which is a extension of Hopfield recurrent neural network, MREM can store not only the binary and bipolar pattern, but also multivalued pattern(such as:red,green,blue). But the capacity of MREM is as much as that of Hopfield neural network.In this paper,the kernelized associative autonomous recurrent neural network,KARN is presented, for extending the capacity of MREM. The kernel function is used to map the low dimension pattern to high dimension pattern and power function is used to compute the capacity of MREM. Compared with MREM, KARN has a main advantage: bigger capacity. Our experimental results manifest this advantage.The paper also proposes multivalued spectral associative autonomous recurrent network, MSARN,which combine the multivalued pattern and spectral associative autonomous recurrent neural network(SARN)[12].Compared with SARN, MSARN has two main advantages: wider value extension and loser interaction.