Memory-based sigma-pi-sigma neural network

In this study, a memory-based sigma-pi-sigma neural network is investigated. The neural network is composed of a different order of submodules, each one consisting of associated single-variable memory arrays. The memory contents in these submodules are adjusted during the learning process. The neural network adds the outputs from these submodules to generate the network output. Each submodule is a memory-based pi-sigma (product-of-sum) neural network. The new structure can learn to implement static mapping that multilayer neural networks (MNNs) and radial basis function networks (RBFs) usually do. Due to the nature of local learning in the associative memory techniques, the learning in the new structure is much easier than that in MNNs. The new neural network structure demonstrates excellent learning convergence characteristics. Using single-variable memory arrays reduces the memory size and overcomes the possible extensive memory requirement problem in RBFs and CMACs in high-dimensional modeling. The simple structure and small memory size make the memory-based sigma-pi-sigma neural networks very attractive.