Abstracts
- 09/01/2012 - V. A. Huynh-Thu
- 09/01/2012 - R. Küffner
- 09/01/2012 - C. Ambroise
- 16/09/2011 - M. Bunster
- 27/04/2011 - P. Meyer
- 06/04/2011 - D. Débois
- 30/03/2011 - J. Wouters
- 28/01/2011 - J.-C. Lambert
- 13/12/2010 - J. Giard
- 17/11/2010 - B. Charloteaux
- 12/05/2010 - P. Soumillion
- 30/04/2010 - J.-P. Vert
- 14/04/2010 - X. Tordoir
- 24/03/2010 - J.C. Voegel
- 17/03/2010 - A. Turtoi
- 24/02/2010 - L. Palmeira
- 27/01/2010 - D. Devos
- 09/12/2009 - F. Heuze
- 25/11/2009 - L. Geris
- 19/11/2009 - V. Gabelica
- 14/10/2009 - G. Zocchi
- 30/09/2009 - K. Van Steen
- 12/06/2009 - D. Gianola
- 05/06/2009 - F. J. Theis
- 02/06/2009 - M. Babu
- 07/05/2009 - J. Eyzaguirre
- 22/04/2009 - E. Bullinger
- 08/04/2009 - F. Francis
- 04/03/2009 - Kick-off
Independent subspace analysis and extraction
Fabian J. Theis
Computational Modeling in Biology - http://cmb.helmholtz-muenchen.de
IBIS, Helmholtz Center Munich & Mathematics, TU Munich
Date and place: Friday June, 5th 11:00 am at 2.93 (B28 - Montefiore)
The increasingly popular independent component analysis (ICA) may only be applied to data following the generative ICA model in order to guarantee algorithm-independent and theoretically valid results. Subspace ICA models generalize the assumption of component independence to independence between groups of components. They are attractive candidates for dimensionality reduction methods, however are currently limited by the assumption of equal group sizes or less general semi-parametric models. By introducing the concept of irreducible independent subspaces or components, we present a generalization to a parameter-free mixture model, and proof separability.
More generally, we ask how to identify and extract subspaces in data based on statistical properties such as non-Gaussianity or signal color (autocorrelations). Algorithmically, we discuss joint block diagonalization with unknown block sizes, and a single-subspace extraction version of this approach. In order to relieve the issue of local minima, we introduce a robust algorithm based on the idea that any ISA algorithm converges to a solution, in which subspaces that are members of the global minimum occur with a higher frequency. This idea allows a blind approach, where no a priori knowledge of subspace sizes is required.