Back-propagation algorithm based on the extended Kalman filter

Recently, Watanabe et al. proposed a back-propagation algorithm, in which the learning rate is time-varying, based on the extended Kalman filter (EKF). In the algorithm the interconnection strengths and biases are treated as the independent variables. However, the interconnection strengths and biases are not always independent, and have generally the mutual correlations. In this paper, we propose a new back-propagation algorithm in the case of considering the mutual correlations between the interconnection strengths and biases. Furthermore, by applying the proposed learning algorithm to the XOR problem and comparing with the algorithm of Watanabe et al., the ability of the proposed learning algorithm is examined.