Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!uunet!pipex!sunic!sunic.sunet.se!news.uni-c.dk!dkuug!dkuug!sandes.dk!m386.pub.ro!mmihai
From: mmihai@m386.pub.ro (Mihai Manolescu)
Subject: Re: Backprop written with integer/fixed pt maths ??
X-Newsreader: TIN [version 1.2 PL2]
Organization: PUB Bucharest (UUCP private)
Message-ID: <D6BCBo.2J5@m386.pub.ro>
References: <D5s0nE.56F@news.sgp.hp.com> <J62avoI.predictor@delphi.com> <3l6b7u$cb@liberator.et.tudelft.nl>
Date: Fri, 31 Mar 1995 16:26:11 GMT
Lines: 37

Richard Kooijman (richard@dutepp6.et.tudelft.nl) wrote:

: My experience with fixed point neural net software showed me that the 
: fixed point version sometimes gets stuck, at least more often than the 
: floating point version. So I think precision does matter.
: I also used 16 bits.

: Of course you might get more stable results by lowering the learning
: rate and such but then you will lose a bit of the training speed
: you were hoping to get in the first place by using a fixed point
: arithmetic.

: If there is enough interest I can place the code on our ftp site.
: It will be available as-is with not much doc. And it is quite old and
: arcane.

: My last neural network software uses floating-point and can deal with 
: recurrent nets. It is already on our ftp site.

hi! i'm interested in public domain source code of fixed point neural nets.
please write me how can i get it through the net.
thanks for your help.

mmihai

-----------------------------------------------------------------------
        Mihai-Costin Manolescu          | address:
 email:                                 | str Arh Stef Burcus, nr 16
        mmihai@pub.ro                   |     Bucharest, Romania
        mmihai@m386.pub.ro              |      tel: +40-(1)-6183167
                                        | fax/data: +40-(1)-6181681
-----------------------------------------------------------------------
-- 

mmihai

-----------------------------------------------------------------------
