home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!decwrl!access.usask.ca!kakwa.ucs.ualberta.ca!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: Reducing Training time vs Generalisation
- Message-ID: <arms.714514725@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <Bt9GIx.9In.1@cs.cmu.edu> <arms.714289771@spedden> <?.714340972@tazdevil>
- Date: Sat, 22 Aug 1992 20:18:45 GMT
- Lines: 41
-
- (Quite a bit of the original has been deleted)
-
- henrik@mpci.llnl.gov (Henrik Klagges) writes:
-
- >Hm, Bill - this reasonability applies to your example problem as well, which
- >is pretty much constructed ad hoc. Just add a training point between your
- >center - or remove the symmetry otherwise, and your extremum is gone (!).
-
- Add a training point and the peak is gone -- true. Changing the
- symmetry does nothing to remove the extremum though.
-
- Pardon a strategic hint: you should avoid weakly pejorative words like
- "ad hoc" in describing the example. You need stronger adjectives.
- Even "irresponsible" or "absurd", which have been tried, failed to
- damage it.
-
- >Concerning lazy evaluation: It is difficult to program on parallel machines.
- >It is a kind of runtime loadunbalancing that constantly switches off tasks
- >(=subtree evaluations) of varying size essentially at random. No way to do
- >that in SIMD, and darn difficult to do it in MIMD at all, and especially not
- >with a high efficiency.
-
- I agree if you only have one task. Then you would try to use eager
- evaluations to get speed, wasting resources. But if you can timeshare
- tasks and only do the necessary ones, you can come out ahead. On the
- Myrias SPS-x machines (MIMD), we have trained several trees at once on
- up to 64 processors. A one-line change to the code was required which
- took 20 minutes first time. However, that machine is perhaps not
- typical of the programming effort required, since it makes certain
- parallel speedups trivial to program. Execution of trained ALNs would
- not have benefitted due to the exceedingly small granularity of the
- task.
-
- Efficiency on the Myrias was about 85%, which was felt to be very
- good.
-
- --
- ***************************************************
- Prof. William W. Armstrong, Computing Science Dept.
- University of Alberta; Edmonton, Alberta, Canada T6G 2H1
- arms@cs.ualberta.ca Tel(403)492 2374 FAX 492 1071
-