A module structured recurrent neural network capable of memorizing and regenerating dynamics

In this report, a module structured recurrent neural network whose size is adaptively determined in a learning process is proposed. The network has the ability to memorize and regenerate any waveforms. In particular, this report shows any periodical waveforms can be approximated by using the minimum number of elementary modules. This network is constructed by adaptive oscillating modules. The adaptive oscillating module consists of two simple neuron nodes. Each node effects the other and itself for oscillating and all weights on connections are adaptively learned. The learning algorithm is based on the modified BP method. The learning of the total network is based on a different criterion called a constructive learning algorithm. In this algorithm, each module can independently learn with suitable speed for given input data. Some simulation examples are demonstrated to check the effectiveness of the proposed network structure and the learning algorithm.