home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!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: What's in a layer ? (was: dumb question on layer enumeration)
- Message-ID: <1992Jul24.143726.50351@cs.cmu.edu>
- Date: 24 Jul 92 14:37:26 GMT
- Article-I.D.: cs.1992Jul24.143726.50351
- Organization: School of Computer Science, Carnegie Mellon
- Lines: 34
- Nntp-Posting-Host: sef-pmax.slisp.cs.cmu.edu
-
-
- From: piccolbo@ghost.dsi.unimi.it (antonio piccolboni)
-
- I'd like to shift your attention to a more compelling problem (for me):
- can we train a neural net with connections of the third type by means
- of standard back-propagation? In our experience this one works as far as
- the connection relation ( (a,b) belong to R iff the weight of connection
- from a to b is different from 0) is a directed acyclic graph.
- Is there a formal proof of this? (I think it would consist in proving
- that the partial derivatives of error with respect to connections weight
- are spatially and timely local).
-
- It's unclear to me exactly what you're trying to prove, but the backprop
- works just fine on networks with shortcuts, as long as there are
- well-defined output units and some total ordering of the units such that
- unit n does not depend on the values of any unit m, where m >= n. The
- usual derivation of the formula for the partials dE/dw does not assume any
- kind of layered structure.
-
- By the way, non-zero weights are not required either. A zero weight will,
- in general, evolve into something more useful. What you don't want to do
- is initialize *all* the weights to zero, or to any other common value.
- This creates symmetries that, in the absence of noise, can't be broken, so
- the net will get stuck.
-
- -- Scott
- ===========================================================================
- Scott E. Fahlman
- School of Computer Science
- Carnegie Mellon University
- 5000 Forbes Avenue
- Pittsburgh, PA 15213
-
- Internet: sef+@cs.cmu.edu
-