home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!munnari.oz.au!metro!news
- From: marwan@metro (Marwan Jabri)
- Subject: Re: Integer neural networks?
- Message-ID: <1992Dec16.235729.23143@ucc.su.OZ.AU>
- Sender: news@ucc.su.OZ.AU
- Nntp-Posting-Host: brutus.ee.su.oz.au
- Organization: Sydney University Computing Service, Sydney, NSW, Australia
- References: <R5BFVB4w165w@west.darkside.com>
- Date: Wed, 16 Dec 1992 23:57:29 GMT
- Lines: 32
-
- Two papers that address the problems of limited precision MLPs in the
- neuroprose archive (anonymous on archive.cis.ohio-state.edu under
- pub/neuroprose) starting with yun....
-
- One paper is on the effects and another paper is on an algorithm tha we
- have developed and successfully used in training limited precision chips.
-
- Although some may argue that limited precision is not a problem, the
- bottom line is if you want to use gradient descent based algorithms you
- will find that backpropagation based techniques will simply fail to do
- anything useful on may applications when the number of bits drop towards
- the 8-6 bits. So the essence is that local search techniques will fail in
- these conditions.
-
- If you have more than 8 bits, you have a good chance in getting gradient
- descent based techniques to work.
-
- On thing you may consider is training (if training can be done off the
- hradware) is use floating point training and then simply scale and
- truncate your values.
-
- If you are interested in an environment that you can use to do limited
- precision training, I have sent some information recently about MUME which
- implements many limited precision algorithms, including that referenced
- above. I can mail you some information if interested.
-
- Marwan Jabri
- marwan@sedal.su.oz.au
- -------------------------------------------------------------------
- Marwan Jabri Email: marwan@ee.su.oz.au
- SEDAL, Electrical Engineering, Tel: (+61-2) 692-2240
- Sydney University, NSW 2006, Australia Fax: (+61-2) 692-3847
-