Using k-fold cross validation in a neural network

Question

Hi,

I'm attempting to implement k-fold cross validation in an functional neural network. The training process is online, and I'm using k=10 for my initial testing.

My question is: what is the best way to use the errors to figure out when to stop training? I compute the root mean squared error for each particular training instance (which gives me around 3200 of these values), but what kind of conditions should be used for stopping?

Is it acceptable to stop half way through one of the k no. of training runs or should it only be done at the end, and what condition tells you that should stop?

Thanks in advance for any help!