steensj@daimi.aau.dk (Steen Sj|gaard) (04/04/91)
Hi,
I am working on a project which deals with different network architectures and
generalization.
What I'm most interested in is to find out if there are any general ways to
determine the connectivity/arrangement of the "necessary" number of hidden
units, when the main subject of interest is the networks' generalization
ability. However, I desperately need a "well-sized" generalization problem
to train and test the different networks on. (By "well-sized" I mean a problem
which definitely is more complex (and realistic) than xor, parity and similar
toy-problems, but on the other hand also less complex/time-consuming than
NetTalk, e.g.)
I have talked to a lot of people about such a problem, but nobody seems to
know of a "standard" or benchmark problem when it comes to analyzing
generalization in neural networks.
As I am sure that I am not the only one who finds this interesting, I would
therefore like to advertise for problems which actually have been successfully
applied to investigate the generalization ability of neural networks.
Any comments, ideas, suggestions, experiences????
Thanks in advance
Steen Sjoegaard
Comp. Sci. Dept.
Aarhus University
DK-8000, Denmark
Email: steensj@daimi.aau.dkegel@neural.dynas.se (Peter Egelberg) (04/11/91)
In article <1991Apr4.130549.12904@daimi.aau.dk> steensj@daimi.aau.dk (Steen Sj|gaard) writes: >Hi, > . . . >I desperately need a "well-sized" generalization problem >to train and test the different networks on. (By "well-sized" I mean a problem >which definitely is more complex (and realistic) than xor, parity and similar >toy-problems, but on the other hand also less complex/time-consuming than >NetTalk, e.g.) . . . Have you tried the two-spiral problem? The objective of the problem is to train the network to predict in which of the two spirals a given point lies in. The input points are given in XY-coordinates. The spirals start out at the center and spiral outwards. One spiral starts out pointing to the left and the other starts out pointing to the right. -- Peter Egelberg E-mail: egel@neural.dynas.se Neural AB Phone: +46 46 11 00 90 Otto Lindbladsv. 5 Fax: +46 46 13 60 85 223 65 LUND, SWEDEN