home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!news.univie.ac.at!chx400!iris-dcp.es!polar.etsiig.uniovi.es!tuya
- Newsgroups: comp.ai.neural-nets
- Subject: Kohonen Algorithm used in ANSIM Neural Net Software
- Message-ID: <1992Nov6.123143.94@polar.etsiig.uniovi.es>
- From: tuya@polar.etsiig.uniovi.es (Javier Tuya)
- Date: 6 Nov 92 12:31:43 +0100
- Organization: Universidad de Oviedo
- Lines: 47
-
- I'm using the ANSim neural net software, with Delta FPP card for some
- simulations in speech recognition tasks.
-
- I've used Kohonen network, an I have found that the algorithm used is not the
- "standard" Kohonen algoritm. Also, I had some results showing that the "ANSim"
- algorithm can be better (in some cases) than "standard" algorithm.
-
-
-
- In "ANSIm" algorithm, the output for each cell j is
-
- out(j) = - SUM(x(i)*m(i,j) + Norm_Weights
- i
-
- The winner is the most negative cell.
-
- The term Norm_Weights is:
-
- Norm_Weights(j) = 0.5 * SUM (m(i,j)*m(i,j))
- i
-
-
- This is different to "standard" algorithm where:
-
- out(j) = SUM(x(i)*m(i,j)
- i
-
-
- Someon knows references (papers or books) where "ANSim" algoritm is used?
-
- I would appreciate also more details about the mathematical
- foundations (ex. convergence proofs), because the classification
- made with "ANSim" algorithm is different that "standard" algorithm.
- For example, "ANSim" algorithm is more sensitive with the magnitude of input
- vectors, due to the additive effect for Norm_Weights.
-
-
- Thanks in advance
-
- +--------------------------------------+------------------------------------+
- | Javier Tuya | PSI: PSI%(02145)285060338::TUYA |
- | E.T.S. Ingenieros Industriales | E-Mail: tuya@etsiig.uniovi.es |
- | Area de Lenguajes y Sistemas | HEPNET: 16515::TUYA |
- | Informaticos (Universidad de Oviedo) | : EASTVC::TUYA (area 16.131) |
- | Carretera de Castiello s/n. | Phone: (..34-8) 518 2049 |
- | E-33394 GIJON/SPAIN | FAX: (..34-8) 533 8538 |
- +--------------------------------------+------------------------------------+
-