At what point does variation in neural network training become unreasonable?

I have a network trained in patternnet. By observing the confusion matrix after each training session, I am witnessing a variance of around 1.5% each time I retrain (From a 2% error rate to a 0% error rate). I feel like there is a big difference in these two percentages. Is there a way to reduce this variation or is it expected with this network?

Answer 1

[Greg Heath]

Since initial weights are random, the variation in results can be considerable. That is why I routinely design 10 to 30 nets for each value of hidden nodes I consider. if you search in BOTH the NEWSGROUP and ANSWERS using

 greg patternnet

you will find many examples. You can reduce the number by trying

 greg patternnet tutorial

Hope this helps.

Thank you for formally accepting my answer

Greg