Page 1 of 1

Cascade Correlation Addition of Hidden Nodes: Won't Work

Posted: Wed 21. Jul 2010, 18:51
by MrCreosote
I have tried CC on a small network that I have successfully run through NeuroShell 2.

I have yet to succeed in forming a single hidden node.

QUESTION: Will the new hidden node be shown in the Network Editor view where my Input and Output nodes are displayed? If NO, how do I tell what is being generated, architecture wise, by the algorithm?

If YES: QUESTION: I have tried varying training settings extensively and have not been able to successfully generate a single hidden node. What could I be doing wrong?

Thanks
Tom

Re: Cascade Correlation Addition of Hidden Nodes: Won't Work

Posted: Wed 21. Jul 2010, 20:03
by Admin
I recently found that there has been a bug introduced that causes the Cascade Correlation Teachers to stop with the message 'Target Net Error reached' before any hidden neurons have been added during the teach process. I've already fixed this and hope to get a new release out within the next days.

Is this what you are observing or are you struggling with other observations?

Regards,
Thomas

Re: Cascade Correlation Addition of Hidden Nodes: Won't Work

Posted: Wed 21. Jul 2010, 20:10
by MrCreosote
YES again Thomas!

I was getting those Target messages just like you said. I then went into a loop of trying to lower and lower thresholds to keep that from happening with no success.

Looks like I was probably OK in that I got the messages which meant the Hiddens should have been created.
_________________________

I have a specific question about the use of CC which I will post in another topic since this one is specific and pretty much closed.

Thanks again,
Tom

PS. When I saw how old the posts were on this forum, I didn't expect such timely response! THanks SO MUCH!

Re: Cascade Correlation Addition of Hidden Nodes: Won't Work

Posted: Wed 21. Jul 2010, 21:41
by Admin
Hi Tom,

I decided to geth this fix out immediately. MemBrain 03.06.02.00 which has this fix included can now be downloaded from the MemBrain homepage.

Regards,
Thomas