Matlab ICA packages available

[Ole Winther]

[Our sincere apologies if you receive multiple copies of this email]

We are happy to announce the availability of three Matlab software
packages for Independent Component Analysis (ICA). 

The algorithms are easy to use with no parameters need to be set by
default. All algorithms come with demonstration scripts. 

The packages contain the following three algorithms:

1. Maximum likelihood (Infomax) - the Bell and Sejnowski 1995 algorithm  
Optimization is performed by an effective second order method.
 

2. Mean Field - Bayesian ICA alg. by Højen-Sørensen, Winther and Hansen
Linear and instantaneous mixing.
Estimation of noise covariance - Gaussian noise model. 
General mixing matrix - quadratic, over- and under-complete
                      - free/positivity constraint estimation.
Variety of source distributions, e.g. 
                      - exponential for positive sources
                      - Gaussian for prob. PCA and factor analysis
                      - bi-Gauss for negative kurtosis sources
                      - heavy tailed for positive kurtosis sources.

 
3. Molgedey and Schuster - The Molgedey-Schuster decorrelation algorithm 
Square mixing matrix and no noise.
Very fast - no iterations.
The delay Tau is estimated.

  
Log likelihoods are calculated for all three algorithm. As an additional
feature the Bayesian Information Criterion (BIC) can be used for
selecting the number of independent components.


The packages are available from:

http://isp.imm.dtu.dk/toolbox/
 
All comments and suggestions are welcome.

Authors:  

Thomas Kolenda     http://isp.imm.dtu.dk/staff/thko/
Ole Winther        http://isp.imm.dtu.dk/staff/winther/
Lars Kai Hansen    http://isp.imm.dtu.dk/staff/lkhansen/

Best regards,

Ole Winther 
-- 
Associate Professor, Digital Signal Processing
Informatics and Mathematical Modelling (IMM)
Technical University of Denmark (DTU)  	http://www.imm.dtu.dk 	
Tel: +45 4525 3895	Homepage: http://isp.imm.dtu.dk/staff/winther/