Neural network has been widely used for nonlinear mapping, time-series estimation and classification. The backprop-agation algorithm is a landmark of network weights training. Although the vast weights update algorithms have been developed, they are often plagued by convergence to poor local optima and low learn velocity. The unscented Kalman filter is a nonlinear parameter estimation algorithm. By means of it, weights update can be realized. Higher training velocity and mapping accuracy of network can be obtained. The numerical simulation results show the effectiveness of the algorithm compared with the standard backpropagation.