by Ingo Schießl, Holger Schöner, Martin Stetter, Anca Dima, Klaus Obermayer
Abstract:
In the separation task of linear mixtures from real experiments the dependencies of the original sources often make “classical” independent component analysis (ICA) algorithms fail. One way to overcome this drawback is the introduction of additional knowledge we have about the mixing process. We introduce a regularization term to the cost function of multishift extended spatial decorrelation (multishift ESD) that punishes the deviation of the time course of the estimated sources from a assumed time course during an experiment. In the case of optical imaging such knowledge can be achieved from the metabolic response of signals to the stimulus onset. We show how the regularization term improves the separation result at different noise levels. The simulations were run on a artificial toy dataset and one dataset that contains prototype signals from a real optical imaging experiment.
Reference:
Regularized Second Order Source Separation (Ingo Schießl, Holger Schöner, Martin Stetter, Anca Dima, Klaus Obermayer), In Proceedings of ICA00 ― International workshop on Idependent Component Analysis and Blind Source Separation (P. Pajunen, J. Karhunen, eds.), volume 2, 2000.
Bibtex Entry:
@inproceedings{schiesl_regularized_2000,
address = {Helsinki, Finland},
title = {Regularized Second Order Source Separation},
volume = {2},
abstract = {In the separation task of linear mixtures from real experiments the dependencies of the original sources often make “classical” independent component analysis ({ICA}) algorithms fail. One way to overcome this drawback is the introduction of additional knowledge we have about the mixing process. We introduce a regularization term to the cost function of multishift extended spatial decorrelation (multishift {ESD}) that punishes the deviation of the time course of the estimated sources from a assumed time course during an experiment. In the case of optical imaging such knowledge can be achieved from the metabolic response of signals to the stimulus onset. We show how the regularization term improves the separation result at different noise levels. The simulations were run on a artificial toy dataset and one dataset that contains prototype signals from a real optical imaging experiment.},
booktitle = {Proceedings of {ICA}00 ― International workshop on Idependent Component Analysis and Blind Source Separation},
author = {Schießl, Ingo and Schöner, Holger and Stetter, Martin and Dima, Anca and Obermayer, Klaus},
editor = {Pajunen, P. and Karhunen, J.},
year = {2000},
keywords = {blind source separation, optical imaging, regularization},
pages = {111--116},
file = {download/publications/Schiessl2000_ICA_RegularizedBSS.ps.gz}
}