Ensemble Joint Approximate Diagonalization by an Information Theoretic Approach
Joint approximate diagonalization (JAD) is a widely-used method for blind source separation, which can separate non-Gaussian sources without any other prior knowledge. In this paper, a new extension of JAD (named ensemble JAD) is proposed in order to ameliorate the robustness for a small size of samples by an information theoretic approach. In JAD, the cumulant matrices are estimated and represented by the average (namely, the first-order moment) over given samples. On the other hand, ensemble JAD preserves the ensemble of all the cumulant matrices for each sample without averaging them. Then, the second-order moments among the ensemble are utilized for estimating the sources. Numerical experiments verify the validity of this method when the sub-Gaussian (negative kurtosis) sources are included.
Keywordsindependent component analysis joint approximate diagonalization information theoretic approach higher-order cumulants
Unable to display preview. Download preview PDF.
- 1.Amari, S., Cichocki, A.: A new learning algorithm for blind signal separation. In: Touretzky, D., Mozer, M., Hasselmo, M. (eds.) Advances in Neural Information Processing Systems 8, pp. 757–763. MIT Press, Cambridge (1996)Google Scholar
- 3.Cardoso, J.F., Souloumiac, A.: Blind beamforming for non Gaussian signals. IEE Proceedings-F 140(6), 362–370 (1993)Google Scholar
- 4.Cichocki, A., Amari, S.: Adaptive Blind Signal and Image Processing: Learning Algorithms and Applications. Wiley (2002)Google Scholar
- 5.Hyvärinen, A., Karhunen, J., Oja, E.: Independent Component Analysis. Wiley (2001)Google Scholar