Using PF to train the neural networks can overcome the drawbacks of BP algorithm such as falling into local minima and slow convergence, and the linearized error of Extended Kalman Filter.