|
Abstract:
This tutorial aims at gathering most of the state of the art in
blind techniques, including the general purpose tool called
Independent Component Analysis (ICA), or more generally, Blind
Source Separation (BSS) from linear or non linear static or
convolutive mixtures. Theoretical concepts are constantly
illustrated by Matlab demos.
Observation models are first surveyed, and inherent indeterminacies
are pointed out. Then, optimization criteria are reviewed. Some
criteria, called contrasts, hold valid in the presence of noise of
unknown statistics. The particular case of semi-blind techniques,
which exploit more information about the inputs, is tackled. In a
second part important results on linear mixtures are presented,
included recent advances concerning methods exploiting
nonstationarity or colorness of sources. In a third part, non linear
mixtures are introduced in a general framework; the particular case
of Post Non Linear (PNL) mixtures is analyzed in depth. The
connection between PNL system equalization and blind inversion of
Wiener systems is eventually established. In a fourth part, some
specific algorithms are presented, including joint diagonalization,
pair sweeping, equivariant adaptive, deflation... Lastly, the
numerous potential applications of ICA are quoted.
Contents
- Introduction.
- Model, separability, and indeterminacies
- Independence criteria and contrast functions
- Semi-blind hypotheses
- Source separation in linear mixtures
- Model and assumptions, static and convolutive cases
- Parameterizations, system of equations
- Sweeping of pairs, Matlab demos
- Whiten or not whiten
- Semi-blind: iid, colored or non stationary sources
- Source separation in nonlinear mixtures
- Constrained nonlinear mixtures
- Example of Post-NonLinear (PNL) mixtures: model and equations
- Matlab demos
- Example of multiplicative mixtures
- Other separable nonlinear mixtures
- From PNL to Wiener systems, matlab demo
- Spotlights on algorithms
- Joint diagonalization (JADE, SOBI, PAJOD)
- Equivariant algorithms (EASY, etc.)
- Deflation (Fast-ICA)
- Algorithms based on mutual information
- Applications
|
|
About the Presenter:
Christian Jutten received the PhD degree in 1981 and the Docteur ès
Sciences degree in 1987 from the Institut
National Polytechnique of Grenoble, France. He taught as associated
professor in Ecole Nationale Supérieure d'Electronique et de
Radio-Électricité of Grenoble from 1982 to 1989. He was visiting
professor in Ecole Polytechnique Fédérale de Lausanne in 1989,
before to be full professor in
Université Joseph Fourier of Grenoble.
For 15 years, his research interests are learning in neural
networks, source separation and Independent Component Analysis.
Since 1988, he is author or co-author of more than 25 papers in the
international journals (Signal Processing, IEEE Trans. on Signal
Processing, etc.) and 70 papers in international conferences. He has
been associate editor for IEEE Trans on Circuits and Systems
(93-95). With J.Hérault and P.Comon, he received in 1991 the award
of the best paper in the journal Signal Processing for his paper on
source separation. In 92-95, he has been the coordinator of the
ESPRIT IV project ELENA. In 1999, he co-organized the first
international workshop on signal separation and independent
component analysis (ICA'99) in Aussois (France, January 1999).
|
