Violation of independence assumption in ICA and its consequences

Independent component analysis (ICA) model assumes the components being separated as statistically independent. The article focuses on the study of violation of this independence assumption by the sources and their consequences in ICA applications, specifically for blind source separation (BSS). It defines and empirically validates the types of violations of independence assumption and their consequences. This brings widening and sharpening of the overlearning phenomena. Overall, the article brings a discussion to consider ICA and BSS both to be different even in linear mixing conditions, as well as a proposal to consider "ICA as BSS" to be a different problem than the usual ICA and BSS problems.