Neural Processing Letters

, Volume 25, Issue 2, pp 91–110

Robust Prewhitening for ICA by Minimizing β-Divergence and Its Application to FastICA

  • Md Nurul Haque Mollah
  • Shinto Eguchi
  • Mihoko Minami
Article

DOI: 10.1007/s11063-006-9023-8

Cite this article as:
Mollah, M.N.H., Eguchi, S. & Minami, M. Neural Process Lett (2007) 25: 91. doi:10.1007/s11063-006-9023-8

Abstract

Many estimation methods for independent component analysis (ICA) requires prewhitening of observed signals. This paper proposes a new method of prewhitening named β-prewhitening by minimizing the empirical β-divergence over the space of all the Gaussian distributions. The value of the tuning parameter β plays the key role in the performance of our current proposal. An attempt is made to propose an adaptive selection procedure for the tuning parameter β for this algorithm. At last, a measure of performance index is proposed for assessing prewhitening procedures. Simulation results show that β-prewhitening efficiently improves the performance over the standard prewhitening when outliers exist; it keeps equal performance otherwise. Performance of the proposed method is compared with the standard prewhitening by both FastICA and our proposed performance index.

Keywords

independent component analysisβ-prewhiteningrobustnessadaptive selectionone standard error

Copyright information

© Springer 2007

Authors and Affiliations

  • Md Nurul Haque Mollah
    • 1
  • Shinto Eguchi
    • 2
  • Mihoko Minami
    • 2
  1. 1.Department of Statistical ScienceThe Graduate University for Advanced StudiesTokyoJapan
  2. 2.The Institute of Statistical MathematicsThe Graduate University for Advanced StudiesTokyoJapan