基于权值函数神经元的BP网络研究

研究了一种神经元模型,在该模型中将参数可调的激励函数往前移到权值上,即把权值变为参数可调的函数,这些权值函数的累加和作为神经元的输出.将此类神经元称为权值函数神经元,根据BP算法给出了由其构成的前馈神经网络的学习算法.仿真实验对比结果表明,在给定的误差精度要求下,基于权值函数神经元的BP神经网络每次训练都能收敛,且平均迭代步数较少,其收敛速度要优于传统BP网络,具有较好的研究应用价值.

A Study on BP Networks Based on Weight-function Neuron

A kind of neuron model is proposed. In the model, the activation function with adjustable parameters is moved forward to the weight, that is the weight becomes a function with adjustable parameters, and the sum of these weight functions as the neuron output. Such neuron is called weight-function neuron. According to BP algorithm, the learning algorithm of feed-forward neural networks with the weight-function neuron is given. Simulation comparison results show that, in a given accuracy requirement, the BP neural network based on the weight-function neurons can converge in each training, and its average iterations is fewer, and its convergence speed is superior to traditional BP network, so it has good research and application value.