INITIALIZED JACOBI OPTIMIZATION IN INDEPENDENT COMPONENT ANALYSIS
Juan José Murillo-Fuentes, Rafael Boloix-Tortosa · 2003
The authors propose a new solution to the minimization of marginal entropies (ME) in multidimensional indepen-dent component analysis (ICA). Starting from the ’Jacobi optimization ’ (JO), we focus on a novel method based on initialization. In this method, we first compute the moment matrix for the prewhitened inputs. Then, the moments at each iteration of this Initialized Jacobi Optimization (IJO) are computed as rotations of this matrix. We include a com-putational comparison between the JO and IJO to design the Optimized Jacobi Optimization (OJO). This new me-thod is available for a wide set of fourth order based con-trasts. Experiments have been included to show that this algorithm have an excellent performance at a low computa-tional cost and memory requirements in comparison to other well-known algorithms. 1.