具有节点偏置的高阶神经网络模型

A HIGHER-ORDER NEURAL NETWORK MODEL WITH NODE BIAS

在汪涛文献基础上提出了一个具有节点偏置的高阶神经网络模型、给出了模型的哈密顿量和学习算法,证明了学习算法的收敛性,该模型能对每一神经元自动引入一个节点偏置使得网络能够存储所有学习图样包括相关图样,其存储容量远高于Hebb—rule—like学习算法下的高阶神经网络模型.对由30个神经元组成的二阶神经网络进行了计算机仿真,结果证实了上述结论.此外,对初始突触强度对学习效果的影响和不同存储图样数目下的平均吸引半径进行了仿真计算并分析了所得结果.

This paper presented a higher -order neural network model with node bias.The Hamilton and the learning algoithm of the model were given. The convergence of the learning algorithm was proved. It introduced automatically a node bias to each neuron so that the new model could store all the learning patterns including coherent patterns, thus the storage capacity of the new model was much higher than that of higher -order model using Hebb -rul - e like learning algorithm. Calculations of computer simulation to a 2nd-order system with 30 neurons were carried out. The results confirmed the above conclusion. The relationship between learning effects and initial synapse weights and the relationship between average attraction radius and number of stored patterns were simulated and analysed. All these remarkable features enable the new model to be of good prospects in application.