home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!darwin.sura.net!mojo.eng.umd.edu!disney.src.umd.edu!tedwards
- From: tedwards@src.umd.edu (Thomas Grant Edwards)
- Subject: Re: neural nets and generalization (was Why not trees?)
- Message-ID: <1992Jul24.195138.27183@src.umd.edu>
- Sender: news@src.umd.edu (C-News)
- Organization: Systems Research Center, Maryversity of Uniland, College Park
- References: <arms.711643374@spedden> <4458@rosie.NeXT.COM> <arms.711990060@spedden>
- Date: Fri, 24 Jul 1992 19:51:38 GMT
- Lines: 30
-
- In article <arms.711990060@spedden> arms@cs.UAlberta.CA (Bill Armstrong) writes:
- >paulking@next.com (Paul King) writes:
- >I agree that discovery of subfeatures is useful, but how much extra
- >interconnection is required -- complete connections to all elements in
- >the preceding layer? If you cut down on the number of connections, you
- >cut the time of executing the net. Go too far -- and you eliminate the
- >possibility of finding subfeatures that are shared.
-
- I'll put forth that finding sub-problems in neural systems is not just
- useful, but required for any significant real-world problem.
- Juergen Schmidhueber said it best, when he talked about real-time
- recurrent learning of a robot to find it's way home from university.
- It could either learn every little microscopic detail of how to get
- from place to place, which would be a moot point anyway since learning
- algorithms would fail with that much which has to be learned, or it could
- break the trip down into important sub-goals such as "open the office door"
- which it could later re-use for opening up any kind of door, "walking
- from bldg A to home" "opening up front door," which would again use the
- "open door" sub-goal it already learned, etc.
-
- The point is, machine-learning people and traditional AI people understand
- the need for knowledge organization, and it is real, and we can't get
- around it in connectionism by just throwing more hidden units at the
- problem. We have to begin to cast knowledge organization into a
- connectionist light. Cascade-Correlation is a first-order example of
- trying to break down goals into sub-goals, but I think we can expand on
- error-correlation as an organizational metric.
-
- -Thomas Edwards
-
-