无监督Eidos表相盒中脑状态人工神经元网络模型参数优化选取

Parameter optimization in Eidos brain-state-in-a-box artificial neural network model

本文研究了无监督表相盒中脑状态(EidosBSB)人工神经元网络模型的参数优化选取问题.通过对模型连接矩阵的特征值进行深入分析,发现网络分类能力取决于有效特征值的平稳性和区别性.提出采用有效特征值均值与其他特征值均值的比,作为参数优化选取的依据,给出了选取最优参数的具体方法.仿真结果表明,经参数优化选取后的EidosBSB模型与原始EidosBSB模型相比,能够获得更强的噪声适应能力,更好的分类性能.参数优化后的网络对噪声污染率为100%的输入样本的平均识别概率达94%以上.

The parameter optimization for the Eidos brain-state-in-a-box（Eidos BSB） artificial neural network model is considered. By an in-depth analysis to the eigenvalues of the model-connected matrix, it can be found that the network＇s classification ability relies on the stability and distinction of the valid eigenvalues. Thereby, a novel parameter optimization technique is proposed, which is based on the ratio of the valid eigenvalues＇ mean to the others. Then, the details of this parameter optimization method are presented. According to the simulation results, this optimized Eidos BSB model is immune to noise and provides better classification results. More than 94% correct classification rate can be attained for the samples with 100% noise contamination rate by employing this optimized neural network.