独立成分分析(ICA)是一种在给出的随机向量中找出统计独立的数据的统计方法,而过完备独立成分分析则是ICA问题中的一类特殊的情形,它要的源信号的数目比观测信号的数目要多。该文提出了一种基于最短路径算法和自然梯度的解决过完备独立成分分析的新算法Turbo-overcomplete。该算法采用了最短路径方法来推断源信号和采用自然梯度的方法来学习基向量,并采用Turbo—overcomplete算法来进行语音信号分离的实验,并把实验结果与现在的一些过完备独立成份分析算法进行了比较。
Independent component analysis(ICA) is a statistical methods for finding statistically independent data within given random vector, ovecomplete ICA is a special case of ICA, which gives more source vectors than observe vectors. This paper introduces a new overcomplete ICA algorithm, named Turbo-overcomplete algorithm, based on shortest path algorithm and nature gradient. Turbo-overcomplete algorithm uses shortest path method to infer source vector and nature gradient to learn the basis vector. It uses Turbo-overcomplete algorithm to separate speech signals and compares the experimental results with current ovecomplete algorithms.