more than one training set for Kohonen
Submitted by khcon on Mon, 06/02/2008 - 22:12
Hi,
I am experimenting with the Kohonen network (OCR) from your article and was wondering how to add a second or even third training set. It seems that after the initial training set you would have to switch to some kind of supervised training for the subsequent sets.
Any thoughts?
KH
more than one set
Maybe I misunderstood the way the Kohonen network works. I thought that you would have to have as many samples in one set as there are output neurons. This way you know when you are done training when you have as many winners as samples. If you add more samples how do you know that the first '3' (as an example for digits) ends up with the same winner as the second one.
Tell me if I am wrong...
KH
more than one set
I believe it would be best to combine both sets into one large training set and present that. Generally when you retrain a neural network with first one training set and then s second, you weaken the first training set in the process. If I want to add patterns to a neural network I generally retrain completly fromt he beginning using one training set that includes everything.