Re: PP vs. ICA

From: Gottfried Helms (helms_at_uni-kassel.de)
Date: 09/23/04 

Date: Thu, 23 Sep 2004 21:58:16 +0200

Am 23.09.04 19:14 schrieb Aleks Jakulin:

> Jay Liu wrote:
>
>>However, from the viewpoint of a problem formulation, two are almost
>>identical. As you know we can formulate ICA and PP in terms of
>>optimization problems. Basis vectors are the optimization variables
>>be determined. Non-gaussianity such as negative entropy can be used
>>a performance index (or objective function) in both ICA and PP.
>
>
> I mostly agree with you. However, even if others too say the same as
> you about the similarity, I personally find the formulations of PP and
> ICA quite different:
>
> PP: a single highly "dependent" multivariate sub-distribution
> ICA: several mutually independent univariate distributions
> reconstructing the whole distribution
>
> Almost everything is a kind of an optimization problem... PP and ICA
> are both usually implemented with similar kinds of loss functions
> (dependence, independence, entropy, etc), but that's as far as the
> similarity goes.
>
You can find a discussion at the very informative website in finland:
which keeps an online tutorial covering a complete description
if the ICA-method.

  http://www.cis.hut.fi/aapo/papers/IJCNN99_tutorialweb/node23.html

  (this is actually the page about the relation to pursuit projection)

Gottfried Helms