Global Optimal ICA and its Application in Brain MEG Data Analysis

Due to its ability to recover the unobserved signals or sources from mixed observations, as well as its ability to analyze the high order statistics of observed signals, ICA has been widely adopted to analyze the brain image data, financial time series, etc. However, most available ICA algorithms are based on gradient descent approach. For non-convex ICA optimization objective function, such algorithms will likely converge to local optimal solution and the most valuable independent components maybe unreachable. In this paper, a new particle swarm optimization (PSO) based global optimal ICA approach is presented to overcome the above problems. Constrained ICA problem is transferred to a constraint free version which can be solved by PSO algorithm efficiently. Applications in the analysis of the magnetoencephalographic recordings (MEG) illustrate the efficiency of the proposed approach