home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!wupost!gumby!destroyer!ubc-cs!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: Learning what COULD be learned
- Message-ID: <arms.711759321@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <1992Jul7.074650.27125@aber.ac.uk> <13uievINN1mp@iraul1.ira.uka.de> <arms.711663417@spedden> <1992Jul21.082035.8898@aber.ac.uk>
- Date: Tue, 21 Jul 1992 22:55:21 GMT
- Lines: 33
-
- dbk@aber.ac.uk (D B Kell) writes:
-
- >In article <arms.711663417@spedden> arms@cs.UAlberta.CA (Bill Armstrong) writes:
- >>The impediment to learning of one output by others that are difficult
- >>or impossible is closely related to the "Why not trees?" question I
- >>am asking. If you use trees, this harmful interaction can't occur.
-
-
- >Please elaborate much more explicitly! **HOW** do the trees flag that
- >some things are not learnable, others are.
-
- If a tree isn't learning the required task, as shown by a lack of
- further improvement, then you can double the size of the tree for that
- particular output and try again. Does this make it clear how the
- technique of independent trees would be used?
-
- Deciding whether a given function can be *learned* is a different
- issue -- complicated considerably by the question "does it generalize
- well? After all, we are allowing contradictory training points as
- well as pattern recognition problems with overlapping
- class-conditional probabilities; so what does it mean to learn?
- Certainly not just being 100% correct on the training data (and
- "overtrained"). Any boolean function is realizable by an ALN; any
- continuous function can be approximated to any desired degree of
- precision without exhorbitant hardware cost. Discontinuous functions
- can also be produced. However, it sounds like an intractable, even
- ill-defined problem to flag things as "unlearnable".
-
- --
- ***************************************************
- 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
-