home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!munnari.oz.au!metro!sunb!retina!len
- From: len@retina.mqcs.mq.oz.au (Len Hamey)
- Subject: Re: Why not trees?(was dumb question on layer enumeration)
- Message-ID: <1992Jul21.044006.2248@mailhost.ocs.mq.edu.au>
- Sender: news@mailhost.ocs.mq.edu.au (Macquarie University News)
- Nntp-Posting-Host: retina.mpce.mq.edu.au
- Organization: Macquarie University, School of Mathematics, Physics, Computing and Electronics
- References: <1992Jul19.045609.89101@cs.cmu.edu> <arms.711643374@spedden>
- Date: Tue, 21 Jul 1992 04:40:06 GMT
- Lines: 37
-
- In article <arms.711643374@spedden> arms@cs.UAlberta.CA (Bill Armstrong) writes:
- >If you trace backwards from any single output node in a feedforward
- >net with any kind of layer structure, you can construct a tree of
- >nodes that will perform exactly the same computation of that output.
- >Some nodes may have to be duplicated in this process. Now suppose
- >that were the hardware structure you had chosen in the first place.
- >Just as a "man cannot serve two masters", wouldn't the learning task
- >be easier for this tree than for the original net, where some nodes
- >are shared between outputs, and have to provide outputs that satisfy
- >two or more elements in net further on? If so, then either the tree
- >learns better, or can be made smaller (probably a good idea to avoid
- >"overtraining").
-
- Yu and Simmons have an interesting paper "Extra Output Biased Learning"
- in which they show that providing additional outputs using the same
- hidden layer can SPEED UP (!) learning by providing hints that the
- hidden layer finds useful. For example, it is quicker to teach a net
- to solve parity and count the number of 1s in the input at the same time
- than to simply teach it to solve parity. This means that if there are
- relationships between the outputs a shared hidden layer can learn more
- quickly than separate hidden layers for each output. Of course, one
- could easily conceive of examples where shared hidden units have trouble:
- when the outputs are totally unrelated to each other.
-
- Yu and Simmons' paper is available in neuroprose. yu.output-biased.ps.
-
-
- Len Hamey
- Lecturer in Computing
- Macquarie University
- NSW 2109 AUSTRALIA
- len@mpce.mq.edu.au
- --
- Leonard G. C. Hamey (Len) len@mpce.mq.edu.au
- Lecturer in Computing (02)805-8978
- Macquarie University
- NSW 2109 AUSTRALIA
-