Decoupled recursive estimation training and trainable degree of feedforward neural networks

A recursive estimation training algorithm of multilayered neural networks (MNNs) with feedforward connections is proposed. The sequential neuron-decoupled extended Kalman filter equations or SNDEKF formulations are derived based on the extended Kalman filter (EKF) algorithm. Except for the advantage of fast convergence, the computational and storage requirements of SNDEKF training are significantly less than those of EKF, and the SNDEKF can naturally be integrated into the parallel structure for the weight vector of each neuron of the network like the conventional backpropagation algorithm. The effectiveness of the training approach is shown through a nonlinear function approximation example