home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!europa.asd.contel.com!darwin.sura.net!Sirius.dfn.de!mailgzrz.TU-Berlin.DE!cs.tu-berlin.de!fauern!lrz-muenchen.de!sun7!t3611ad
- From: t3611ad@sun7.LRZ-Muenchen.DE (Xinzhi Li)
- Subject: Re: Kohonen nets with other than rectangular output arrays
- Message-ID: <t3611ad.725020473@sun7>
- Sender: news@news.lrz-muenchen.de (Mr. News)
- Organization: Leibniz-Rechenzentrum, Muenchen (Germany)
- References: <phi.724924574@wambenger>
- Distribution: lrz
- Date: Tue, 22 Dec 1992 10:34:33 GMT
- Lines: 23
-
- To the question:
-
- >What about other neighbourhood graphs?
- e.g. toroidal, trees etc.
-
- If you just want a quanzation of you data set, the neighborhood
- graph (network topology) is, by my experience, rather irrelevant. If
- you want to preserve the topological information of the data set in the
- network, the toroidal network topology is more favorable than the
- rectangular topology, since it is boundaryless, all its neurons are
- treated equally during the training process and hence more homogeneous.
-
- The following paper presented a type of Kohonen net with adaptive tree
- topology which seems to be very flexible and general. It is, however,
- much more difficult to implement than the rectangluar or toroidal
- network.
-
- J.A. Kangas, T.K. Kohonen and J.T. Laaksonen: Variant of
- Self-Organizing Maps. IEEE Trans. on Neural Networks, vol.1. No.1 March
- 1990. pp. 93-99.
-
- enar.eros.chemie.tu-muenchen.de
- X. Li
-