home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!gumby!destroyer!ubc-cs!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: Characterizing learnable functions
- Message-ID: <arms.713640420@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <1992Aug10.223138.25927@cco.caltech.edu> <1992Aug11.111206.25386@cs.tu-berlin.de> <arms.713550511@spedden> <1992Aug12.112845.1060@cs.tu-berlin.de> <TMB.92Aug12171333@arolla.idiap.ch>
- Date: Wed, 12 Aug 1992 17:27:00 GMT
- Lines: 42
-
- tmb@arolla.idiap.ch (Thomas M. Breuel) writes:
-
- >In article <1992Aug12.112845.1060@cs.tu-berlin.de> async@opal.cs.tu-berlin.de (Stefan M. Rueger) writes:
-
- > >Maybe consumers of NNs should ask for and insist on Lipschitz conditions:
- > > |output(a) - output(b)| <= const * | a - b |
- > >
- > >for all a, b in the domain of the function.
-
- > These Lipschitz-functions **are** continuous and thus covered by the
- > results of Cybernko, described in my original article.
-
- >Unfortunately, imposing a Lipschitz condition is not enough. While a
- >neural network might be able to approximate any Lipschitz function,
- >the number of training examples (the "sample complexity") you need for
- >actually finding a "good approximation" is provably too large in many
- >cases.
-
- Excellent point. If the training or test data is very sparse, you
- would have to have an infinitesimally small Lipschitz constant to
- specify what you want on the whole space or to determine that the
- result of training is within spec.
-
- On the other hand, if you look at the structure of a net and you try
- to compute a Lipschitz bound at a given point, you would probably have
- to use absolute values of weights, and make pessimistic estimates of
- the derivatives of the sigmoids, so you would end up with a uselessly
- large bound.
-
- I wouldn't give up on the idea of Lipschitz bounds yet, but it looks
- bleak. The only alternative I can see is to assume piecewise
- monotonicity of the desired function in the specification, and to
- force piecewise monotonicity of the result of training.
-
- AN IMPORTANT QUESTION: Can anyone think of any other straightforward
- ways to get bounds on values *between* test points? The application
- of NNs in safety-critical areas is at stake.
- --
- ***************************************************
- 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
-