Improved Multiplicative Orthogonal-Group Based ICA for Separating Mixed Sub-Gaussian and Super-Gaussian Sources
Yalan Ye, Zhilin Zhang, Shaozhi Wu, Xiaobin Zhou · 2006
Recently, the fully-multiplicative orthogonal-group ICA (OgICA) neural algorithm has been proposed, which exploits the known principle of diagonalisation of a tensor of a warped network's outputs. Unfortunately, the algorithm is only able to separate sub-Gaussian source signals. To address this problem, the paper proposes an improved algorithm that adopts two nonlinearities and a flexible nonlinear model switching technique. The improved OgICA algorithm can instantaneously separate not only the mixture of pure sub-Gaussian source signals, but also the mixture of super-Gaussian and sub-Gaussian source signals. Besides, the algorithm has fast convergence speed and high separation performance. The validity and effectiveness of our proposed algorithm are confirmed through extensive computer simulations