Temporal Decorrelation as Preprocessing for Linear and Post-nonlinear ICA

Abstract. We present a straightforward way to use temporal decorrelation as preprocessing in linear and post-nonlinear independent component analysis (ICA) with higher order statistics (HOS). Contrary to the separation methods using second order statistics (SOS), the proposed method can be applied when the sources have similar temporal structure. The main idea is that componentwise decorrelation increases nonGaussianity and therefore makes it easier to separate sources with HOS ICA. Conceptually, the non-Gaussianizing filtering matches very well with the Gaussianization used to cancel the post-nonlinear distortions. Examples demonstrating the consistent improvement in the separation quality are provided for the both linear and post-linear cases.

• Type: Conference
• Year: 2004
• Where: ICA
• Authors: Juha Karvanen, Toshihisa Tanaka
• Keywords: ICA 2004, Order Statistics, Post-nonlinear Independent Component, Similar Temporal Structure