Problems with big nets

[Emil Kaae Hansen]

Dear all, I hope some of you can help me with this problem.

I've got a funny situation with JavaNNS, when I work with large networks. I
am trying to train a RBFN with 198 input/output units, and 200 hidden units.
I've got approx. 30 datapoints (so far) in training and validation sets
respectively.

The problem occurs when I try to train the network. After what seems like a
couple of iterations, the activation of all the hidden units are NaN, and
thus the network stops learning. Is this a JavaNNS problem, or am I doing
something wrong? The network is arranged as this: input units are the first
1x198 units, then 10x20 hidden units, then 1x198 output units. They are
connected in feed-forward style.

And another thing: I have this strange problem initialising the network. If I
load a pattern, and press init in the control panel, I get a message
"SNNS-Kernel Error: Algorithm needs patterns. Please press TEST first to
check patterns.". I can trick JavaNNS into accepting, by pressing the test
button in the Analyse window (with learning enabled, naturally). After this I
can initialise the network. Also, I seem to have trouble selecting different
training and validation sets.

I have not observed these oddities with the included examples.

I would appreciate any help or guidance you can give.

Best regards,
Emil Hansen