home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!sun-barr!cs.utexas.edu!uwm.edu!ogicse!das-news.harvard.edu!cantaloupe.srv.cs.cmu.edu!crabapple.srv.cs.cmu.edu!news
- From: sef@sef-pmax.slisp.cs.cmu.edu
- Newsgroups: comp.ai.neural-nets
- Subject: Re: any complicated units/connections used ?
- Message-ID: <1992Jul27.142202.290451@cs.cmu.edu>
- Date: 27 Jul 92 14:22:02 GMT
- Article-I.D.: cs.1992Jul27.142202.290451
- Organization: School of Computer Science, Carnegie Mellon
- Lines: 38
- Nntp-Posting-Host: sef-pmax.slisp.cs.cmu.edu
-
-
- From: prechelt@i41s14.ira.uka.de (Lutz Prechelt)
-
- in almost all neural network models used today, the connections are
- restricted to multipying their input with a weight and the units
- are restricted to applying a mapping function (e.g. threshold or
- squashing function) to their cumulated inputs.
-
- I would like to know about all *more complicated* forms of units or
- connections anybody is using or planning to use in a neural network.
- I am aware of sigma-pi units, activation decay, and time delay
- connections but please send me info about how you use these, anyway.
-
- The key question to ask about nets with more complex elements is whether
- the added complexity buys you anything important. For example, could each
- of the more complex units be simulated by a little cluster of backprop
- units.
-
- Even if there's some formal equivalence, the more complex units may be
- useful if they introduce some constraint or structure into the net that was
- not there in backprop. This may lead to faster training or better
- generalization. For example, sigma-pi units are not fundamentally more
- powerful than plain old sigma units with an extra layer or two, but they
- seem to be well-matched to problems that want to activate different rule
- sets (or overlays on a fixed rule set) at different times. For example,
- such units have been used to good advantage in speech understanding to
- blend together several different speaker models. (See paper by Hampshire
- and Waibel in NIPS-2.)
-
- -- Scott
- ===========================================================================
- Scott E. Fahlman
- School of Computer Science
- Carnegie Mellon University
- 5000 Forbes Avenue
- Pittsburgh, PA 15213
-
- Internet: sef+@cs.cmu.edu
-