home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!dtix!darwin.sura.net!mips!zaphod.mps.ohio-state.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: Rule extraction from NN
- Message-ID: <1992Jul27.141356.281701@cs.cmu.edu>
- Date: 27 Jul 92 14:13:56 GMT
- Organization: School of Computer Science, Carnegie Mellon
- Lines: 33
- Nntp-Posting-Host: sef-pmax.slisp.cs.cmu.edu
-
-
- The general problem with rule extraction from neural nets is not solvable.
- A neural net provides a richer language than can be captured in a small,
- human-readable set of rules, whether symbolic or fuzzy. Sometimes you can
- look at a trained net and say something like, "The output is the same as
- input B unless both of C and D are present." But more often you're faced
- with, "The output is 3.257 times the sigmoid of B unless C is more than
- 457 times the sigmoid of .698 D plus E time 1.035e-11..." Even if you
- translate into the language of fuzzy rules ("much bigger than" and so on),
- the latter kind of expression is going to be ugly.
-
- In such cases, you can either give a clean, simple rule that is a crude
- approximation to what the network is really doing (but perhaps sufficient
- for your needs), or can extract an expression that gives the whole truth,
- but that is not human-readable in any useful sense. Shavlik and others
- have claimed that in certain restricted domains, with certain restricted
- architectures, almost all the nets produced are easily convertible into
- reasonably simple rules. Personally, I suspect that you can only count on
- this if the restrictions are very severe.
-
- Looked at from the other side, this is a kind of advantage: neural nets
- allow us to explore the wide universe of mappings for which there is no
- simple rule-based equivalent.
-
- -- Scott
- ===========================================================================
- Scott E. Fahlman
- School of Computer Science
- Carnegie Mellon University
- 5000 Forbes Avenue
- Pittsburgh, PA 15213
-
- Internet: sef+@cs.cmu.edu
-