Adaptive Constrained ICAMGGMM: An Improvement Over ICA

Blind source separation is one of the most common methods used to separate mixed signals. The paper proposes a novel algorithm known as acICAMGGMM, which stands for adaptive constrained independent component analysis with a multivariate generalized Gaussian mixture model. This algorithm is based on independent component analysis (ICA) and is designed to separate mixed signals. The acICAMGGMM model relaxes the ICA limitations in multivariate data by considering the second and higher-order statistics, which makes it more flexible to fit different shapes of data. This model takes into account the correlated features in the multivariate data using the full covariance matrix in its cost function. The acICAMGGMM algorithm effectively separates the estimated sources from the original sources flexibly and adaptively. It avoids the use of imprecise constraints that are typically imposed in conventional constrained ICA. The paper validates the proposed model using speech and EEG applications and demonstrates the model's effectiveness compared to base models utilizing a variation of separation metrics. Furthermore, the paper highlights the robustness of acICAMGGMM by assessing its performance using spatial correlation and inter-signal-interference as benchmarks when compared to the baseline models.