A parallel independent component analysis algorithm

Independent component analysis (ICA), orienting as an efficient approach to the blind source separation (BSS) problem, searches for a linear or nonlinear transformation that minimizes the statistical dependence between source signals. However, ICA has been very time consuming in real-time application, especially for high volume data set. In this paper, a SPMD-structured parallel ICA (pICA) algorithm is presented. pICA is developed based on the FastICA approach and conducted in three stages: the estimation of weight matrix in which sub-processes are executed on multiple processors in parallel, the internal decorrelation that performs weight vector decorrelations within the same sub-matrix, and the external decorrelation that performs weight vector decorrelations between different sub-matrices. We propose a LogP-based performance prediction model that estimates the speedup of the pICA process by taking into account the size of the dataset, the network bandwidth, and the processor overhead. We further implement the pICA algorithm in an MPI environment consisting of 10 processors. Both analytical and experimental studies show that pICA distributes the computation burden to multiple processors without losing accuracy. Comparing to FastICA, the pICA process generates an exponential speedup when the number of the estimated weight vectors increases linearly