home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!mcsun!news.funet.fi!network.jyu.fi!hovila
- From: hovila@jyu.fi (Ari Hovila)
- Subject: Adjusting weights in backprop simulators
- Message-ID: <1992Sep9.094735.15604@jyu.fi>
- Organization: University of Jyvaskyla, Finland
- Date: Wed, 9 Sep 1992 09:47:35 GMT
- Lines: 43
-
- In article 6499 of comp.ai.neural-nets jjb@sequent.com (Jeff Berkowitz)
- writes:
-
- (...some stuff deleted)
- >Now, the question. While I was trying to debug my backprop simulator,
- >my wife discovered what appears to be a subtle difference between the
- >precise description of backprop and several of the "C" implementations
- >I've picked up via ftp.
-
- >In the "original" paper (Rumelhart, Hinton, Williams, "Learning
- >Internal Representations by Error Propogation", 1986) the backward
- >pass is described as follows:
-
- > ...The first step is to compute delta for each of
- > the output units. [...] We can then compute the
- > weight changes for all connections that feed into
- > the final layer. AFTER this is done, then compute
- > deltas for all units in the penultimate layer...[etc,
- > emphasis mine.]
- ...
- >as I read the description, I should change the weight FIRST as I back
- >up, and then use the NEW value in the accumulated error. At least
- >the Dayhoff description pretty much states this in black and white.
-
- Well, this has been bugging me too. I have impelented a bp-simulator
- with C, and I use the method you suspected was 'an implementation
- error' (i.e. calculate deltas first and then adjust the weights).
-
- I'm really not sure what the main difference between these
- two approaches is, but at least in a book by James Freeman (Neural
- Networks, algorithms, applications and programming techniques) it
- is very clearly stated that you should calculate deltas for hidden
- layer nodes before adjusting the outgoing weights.
-
- Since weight changes are usually small it could be possible that the
- algorithm works both ways. I doubt that anyone has seen a case where
- one of these methods works while the other one fails...
-
- It would be nice to see other comments on this.
-
- Ari Hovila,
- University of Jyvaskyla, Finland
- internet e-mail address: hovila@jyu.fi
-