摘要
本文把同伦论中零点路径跟踪的概念推广到多层前向网络能量函数极小点路径的跟踪，从而提出多层前向网络一种新的训练方法──同伦ＢＰ算法（包括教师同伦和输入同伦），并且分析了其收敛性质．结合异或问题和由部分信息重建完整信号所作仿真计算，证明此方法在收敛速度和避免陷入局部极小的能力上均明显地优于常规ＢＰ算法，体现了人类认识事物时所遵循的由简到繁、循序渐进的原则．
The concept of zero-path tracking in homotopy is generalized into minimum (of energy function)-path tracking and a new method for the training of multi-layer perceptron is thus put forward──the homotopic BP algorithm. An analysis is made on the convergence property of the proposed method. By computer simulations, it is shown that the speed of convergence and the ability of getting rid of local minimum of the new method are far better than that of the conventional BP method. The rule of learning step-by-step and from simple to complex is reflected in the proposed method.
出处
《计算机学报》
EI
CSCD
北大核心
1996年第9期687-694,共8页
Chinese Journal of Computers
关键词
神经网络
多层感知机
同伦论
BP算法
Artificial neural network, multi-layer perceptron, homotopy, cognitive learning
