Date: Fri, 11 Dec 1998 17:59:20 +0100 (MET)
From: Jean-Francois Cardoso 
Subject: new version of the announcement


Dear Colleague,

Following a suggestion of Simon Haykin, several ICA/BSS researchers
have convened at the NIPS*98 meeting and have agreed on the
opportunity of sharing on the Internet some tools regarding ICA
research.  

The following propositions have emerged.

 1) Creation of a mailing list devoted to ICA/BSS research.

	The list will be initialized with the e-mail addresses of the
	attendees of ICA '99.


 2) Creation and maintenance of an ICA home page containing:

	- links to data sets appropriate for ICA
	- source code for ICA algorithms
	- archive of the ICA mailing list


 3) Evaluation and comparison of ICA/BSS algorithms.

	- This would be a collaborative project to assess the
	  performance and the limitations of existing ICA algorithms
	  on both real and synthetic data sets.

	- A workshop at NIPS (and possibly a book) could be prepared,
	  based on the results of these evaluations.


All people involved in ICA/BSS research are invited to contribute and
to subscribe to the mailing list (directions for doing so will be
distributed shortly).

More details about these projects appear below.  All the points are
open to discussion and we are certainly waiting for your feedback.

Best regards !


------------
This report has been prepared by J. Principe and J.-F. Cardoso.
December  9, 1998



APPENDIX: More details on the items listed above.
=================================================


1) DATABASE of SIGNALS
======================

We will try to collect several significant data sets, both real and
synthetic.

We have agreed on the following rules:

 - The data files will reside on the computer of the provider, and a
   link should be sent to Jean-Francois.  He will keep the links in
   the ICA home page.

 - Each provider will include a brief explanation of the data.

 - Each provider will include a few lines of commented MATLAB code to
   read the data.  This solution is preferred to trying to enforce a
   common format.

We will start collecting the contributions as soon as possible.  If
convenient, a CDROM will be made available.

Several people present at the NIPS meeting have already agreed on
providing the following data sets.  More data sets are welcome.

 Synthetic data
 --------------

 o A selection of synthetic signals (both hard and easy mixtures of
   iid sequences, super- and sub-Gaussian and multimodal
   distributions.  To be prepared by John Fisher (MIT).

 o Mystery datasets (from Terry.  Other contenders ???)



 Real data
 ---------

 o Biomedical data :

 	o FMRI (from Terry Sejnowsky); 

 	o EMG (from Erkki Oja); 

 	o EKG (from Lieven Delathauwer, if permission is granted);

 	o EEG (from Tewon Lee);

 o Digital Communications (from JF Cardoso, if permission to publish
   is granted);

 o Images (from Aapo Hyvarinen, other ?);

 o Speech, computer mixed and real room (from Jose Principe, see also
   Daniel Schobben);

 o Financial (from Andrew Back).




DATABASE of ALGORITHMS
======================

- Researchers are encouraged to contribute their algorithms.
  Submission guidelines will be posted on our site.

- The preferred coding language is Matlab but any language is
  acceptable.

- For instantaneous mixtures, the contributors are encouraged to adopt
  the following input/output model to simplify running algorithms:

 	function B = this_ica_algo(X,m)

  where:
 - X is the n x T data matrix (n sensors, T samples)
 - m is the number of independent components to be extracted
   from the data set.  It should preferably be optional: if
   omitted, it can default to n or be determined by determined
   by the algorithm by an order selection procedure.
 - B is a m x n separating matrix such that B*X is an m x T
   matrix of independent components/source signals.



PERFORMANCE EVALUATION
======================

  There has been quite a lot of discussion regarding performance
  evaluation.  Indeed, making it possible/easy to compare algorithms
  is one of the objectives of the repository.

  In the perspective of organizing performance comparison, the
  following has been proposed: 

  At first we will request the algorithm provider to do the testing of
  his/her own algorithm and post the results in the site according to
  the following guidelines:

  - Algorithms should be compared according to the following
    parameters:

 - Performance measure in Amari's NIPS 96 paper SNR in dB (when
   mixing matrix is available)

 - The output should include the weight matrix and the
   estimation of the mixing matrix to allow listening/visual
   assessment.

 - We should also use listening or visual tests for assessing quality.
   A ranking of the algorithms through a vote (scale of 1 to 10 with
   10 being the highest quality) will be requested from each
   participant. The votes will be compiled.

 - We should characterize the performance of the algorithms in
   terms of memory usage, flops, and the computational
   complexity of the algorithm.

  We should provide the code and the parameters utilized in each test
  in MATLAB. At a later date, the algorithms and parameters will be
  made available in a form that anyone can use them.

  The results should be submitted by May 1, 1999. A workshop at NIPS
  (and possibly a book) will be prepared with the results of the
  submissions.