Improvement to Geometric Linearization of gpICA Algortihm by Compensation and Multiple Points

Abstract

In this paper we show two modifications of the gpICA (geometric post non-linear independent component analysis) algorithm. gpICA algorithm is a novel method to solve the PNL (post non-linear) scheme. We propose these modifications to improve the mean squared error, the correlation of the recovered signals and algorithm reliability. The first improvement, called compensation, takes advantage from the implicit information given by the point to be linearized. On the other hand, while the original gpICA algorithm uses two sets of two points to make an update, our second modification uses two sets of four points. We present experimental results which validates the effectiveness of each modification.