home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!dtix!darwin.sura.net!mojo.eng.umd.edu!disney.src.umd.edu!tedwards
- From: tedwards@src.umd.edu (Thomas Grant Edwards)
- Newsgroups: comp.ai.neural-nets
- Subject: Re: ALN vs BP
- Message-ID: <1992Jul24.194122.27019@src.umd.edu>
- Date: 24 Jul 92 19:41:22 GMT
- References: <1992Jul24.053623.22636@cs.yale.edu>
- Sender: news@src.umd.edu (C-News)
- Organization: Systems Research Center, Maryversity of Uniland, College Park
- Lines: 31
-
- In article <1992Jul24.053623.22636@cs.yale.edu> tsioutsias-dimitris@CS.YALE.EDU (Dimitris Tsioutsias) writes:
- >It seems that after the backprop fans, we have now the ALN ones. Why
- >each group (or any other that shows strong support) is trying to pass
- >its nets as the dominant ones?
-
- Yeah, but there are alot of people who want to throw nets at real problems
- today, and it would be a little silly for anyone to expect gradient
- descent MLP nets at any real problem and expect results.
-
- Clearly, ALN's do the job (learning) real fast. So do locally receptive
- fields (infact, if you have a good grasp on the chaotic dynamics of
- the problem, using self-adaptive receptive fields is a big win).
- Still, even a good conjugate-gradient MLP program should work OK, but I
- don't feel like writing one! Cascade-Correlation is particularly neat,
- especially for really difficult problems.
-
- The thing to remember is that NEURAL NETS DO NOT PERFORM MIRACLES.
- They will not predict the stock market well, nor will they predict daily
- solar flux from the sun months in advance, because there is too much
- missing information besides whatever input you give the net to ever solve
- the problem. They are best used in situations where a person, if he/she
- sat down for a few weeks and worked on the problem given the input you
- give the net, and figure it out. Sonar problems already solved by
- sonarmen, truck backing up problems already figured out by truck drivers,
- etc. They are best utilized as quick-learning first-order solutions to
- problems.
-
- Of course, there is plenty of machine-learning theory to be worked
- out from neural nets, but that is research, not application...
-
- -Thomas Edwards
-