home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.ai.neural-nets:4752 sci.math.stat:2729
- Path: sparky!uunet!cs.utexas.edu!qt.cs.utexas.edu!yale.edu!ira.uka.de!news.belwue.de!iptc!news.wsi!peanuts!goeppert
- From: goeppert@peanuts.informatik.uni-tuebingen.de (Josef Goeppert)
- Newsgroups: comp.ai.neural-nets,sci.math.stat
- Subject: Separation on non-linear and non-convolutive mixtures
- Date: 8 Jan 1993 12:38:07 GMT
- Organization: "Lehrstuhl fuer Technische Informatik, Uni Tuebingen"
- Lines: 40
- Sender: goeppert@peanuts (Josef Goeppert)
- Distribution: world
- Message-ID: <1ijsjfINN477@peanuts.informatik.uni-tuebingen.de>
- References: <1ijn70INNsqf@gap.caltech.edu>
- NNTP-Posting-Host: charlie.informatik.uni-tuebingen.de
- Keywords: artificial neural networks, separation of sources
-
-
- Hello,
-
- I am working in the domain of multi-component analysis of
- sensory output signals. My problem is to seperate different
- influences by observation of sensory output. Unfortunately,
- the Transfer Function of these sensors is non-linear (sine-like).
- The exact nonlinearity of the sensors is unknown and
- varies frome one sensor to another. I am trying to solve the
- problem by neural nets, using supervised Backpropagation
- Net or neural aproches of clustering methods (like Kohonen's SOM),
- it would be better to use smaller and non-supervised methods.
-
- Linear approaches like Principal Component Analysis (PCA)
- and Independ Component Analysis (INCA) or the neural
- inplementation of PCA (Oja) and INCA (Jutten/Herault)
- are not suitable because of big non-linearities.
-
- In order to come to smaller size and non-supervised
- training, I am locking for other architecture or for
- mathematical principles for the separation of sources
- in non-linear and non-convolutive case. Highly redundant
- input-data (in my case hundreds components to
- seperate less than ten sources), may allow nevertheless
- the differentiation and the estimation of the
- non-linearities. Kohonens Self-Organizing-Map turned
- out to do so, but with huge number of neurons.
-
- I would be happy for ideas and pointes to neural
- approaches or statistical methods for this problem.
-
- Thanks in advance
-
- --------------------------------------------------------------------------
- Josef Goeppert | Tel: (49) 7071 29 5940
- Wilhelm Schickard Institut | Fax: (49) 7071 29 5958
- University of Tuebingen | E-mail: goeppert@peanuts.
- Sand 13, D-W7400 Tuebingen, Germany | informatik.uni-tuebingen.de
- --------------------------------------------------------------------------
-
-