规范化自然梯度ICA算法

Standardized Natural Gradient ICA Algorithm

对基于李群不变性的自然梯度ICA算法进行了改进,提出了一种规范化自然梯度ICA算法.该算法通过引入规范化因子,保证参数矩阵的行列式的绝对值在学习过程中恒为1,避免了参数矩阵剧烈变化,使得学习过程更稳定更快速,这种改进还起到简化目标函数的作用,使得规范化自然梯度ICA算法更加简单便利.在BSS模拟实验中,把常规梯度的ICA算法、自然梯度ICA算法与规范化自然梯度ICA算法进行比较,结果表明新算法的信号恢复精度更高,收敛速度更快.

This paper presents a standardized natural gradient ICA （Independent Component Analysis） algorithm through improving the natural gradient ICA algorithm based on Lie Group Invariance. T e new algo th introduces a standardizing factor which makes the absolute value of the determinant of parameter matrix equal to one. Therefore, the learning process is more stable and faster by restricting the drastic change of parameter matrix. In addition, the new algorithm is simpler by using standardizing factor to simplify the general criterion function. In BSS （Blind Signal Separation） simulation experiment, we have compaied three algorithms including the general gradient ICA algorithm, the natural gradient ICA algorithm and the new algorithm. The results show the third is the best in the precision of restoring signals and converges fastest.