Extended Kalman filter training of neural networks on a SIMD parallel machine (Q697007)

The paper presents the design and realization of a multistream decoupled extended Kalman filter (DEKF) for training neural networks. With a large training data set, a large network configuration and multiple network inputs, one needs a powerful computer to execute training. In the paper, a brief characterization of the so-called MasPar SIMD parallel machine and its capacities is given. The algorithm itself, chiefly consisting of five steps for each batch of training data, is presented and its performance on the MasPar parallel machine is studied. An example is given to illustrate the effectiveness and significance of the multistream DEKF training. A comparison is provided between multistream DEKF and batch-mode backpropagation trainings. Besides, one compares the approximate run time needed for network training on parallel and sequential machines. Simulation results of wind power estimation by neural networks are exhibited.