加快神经网络学习的梯度幂次法

An Accelerated Learning Algorithm for Neural Networks by Gradient Power Method

分析了前向神经网络极值点附近的性态，指出基本ＢＰ算法用于分类问题时收敛缓慢的原因．我们利用梯度模的幂次去修改学习率，仿真结果表明，将此方法用于分类问题的训练时，收敛速度明显优于基本的ＢＰ算法．

The property of the extreme points of the multilayer neural networks is studied and the reason for the slow convergence rate of the standard BP algorithm in the vicinity of the extreme point is given. This paper uses the power of the gradient norm to modify the learning rate. Simulations indicate it has a superior convergnce rate for the classification problems, compared to the standard BP algorithm.