home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.ai:5026 comp.ai.neural-nets:4982
- Newsgroups: comp.ai,comp.ai.neural-nets
- Path: sparky!uunet!vuse.vanderbilt.edu!cantwell
- From: cantwell@vuse.vanderbilt.edu (Jeff Cantwell)
- Subject: Wanted: Heuristics for Neural Net Design and Use
- Message-ID: <C1J2BA.Mz9@vuse.vanderbilt.edu>
- Sender: news@vuse.vanderbilt.edu
- Nntp-Posting-Host: necs
- Organization: Vanderbilt University School of Engineering, Nashville, TN, USA
- Date: Wed, 27 Jan 1993 19:33:10 GMT
- Lines: 15
-
- I am looking for some heuristics for practical use of neural networks.
- For instance, how should momentum, slope, magnitude, step size, and
- other parameters be varied if the network fails to converge? How do
- you decide how many hidden units are needed? Are there generally
- applicable rules, or is everything "black magic"? Any pointers to
- articles or papers would be greatly appreciated.
-
- -Jeff
-
-
- --
- Jeff Cantwell, Vanderbilt U. CIS Box 1804, Sta. B Nashville, TN 37235
- Internet: cantwell@vuse.vanderbilt.edu Bitnet: CANTWEJR@VUCTRVAX
- Phone: (615) 322-0110 Fax: (615) 343-6449
- 2 Cor. 12:9 "Power is perfected in weakness."
-