基于典型相关分析的独立分量分析对照函数构建方法研究

Study of Contrast Function Construction on Independent Component Analysis Based on Canonical Correlation Analysis

本文根据典型相关分析的特征,并结合近年来的研究热点———核学习的有关理论,提出了一种在可再生核Hilbert空间为独立分量分析构建对照函数的新方法,并证明其与以前提出的普通对照函数一样,具备统计相关测度函数所需的满意数学特征.通过对各种源分布的分离结果仿真表明,该方法比现有的其他方法具有更好的鲁棒性和灵活性.

By combining the relational theories of kernel learning, which is currently hot problem of machine learning in recent years, with some features of canonical correlation analysis, this contribution suggests new methods of constructing contrast functions for independent component analysis in reproducing kernel Hilbert space. It is justified the contrast function is with all expected attributes of statistical metric about dependence just as same as normal contrast function presented before. The results of simulation, operated on various probability functions of signal source distribution, prove this method to outperform many of the presently known algorithms in robust and flexible properties.