一种基于动力学模拟的BP神经网络快速算法

new fast neural network Back Propagation Algorithm, Multi-Degree-of-Freedom Backpropagation (MDFBP) was proposed to improve the convergence rate of Back propagation Momentum Algorithm. The underlying idea is to consider the process for weight coefficients of neural networks as a particle movement on the MSE supersurface in the gravitational field; by using the Tailor’s approximation, the differential equations can be simplified as a linear Multi-Degree-of-Freedom system at each time; the stability analysis of the MDFBP algorithm is presented and the parameters are determined according the linear system theory based on the stability condition. The simulation results show that the initial convergence rate of MDFBP is much faster than that of the Back propagation Momentum algorithm. The significance of the paper is that it point out that a multi-layer neural network is equivalent to a time-varying lin-ear quadratic function..