You cannot over-train unless the net is over-fit.
The net is over-fit if there are more unknown parameters than there are training equations.
If you accept all default parameters except No. of hidden nodes, H, then the number of unknown parameters for a single hidden layer net with I-H-O node topology is just the number of unknown weights which, for I-dimensional "i"nput vectors and O-dimensional "o"utput target vectors, is
Nw = (I+1)*H+(H+1)*O
With N pairs of input/target training vectors, the default number used for training is
Ntrn = N - 2*round(0.15*N) % ~0.7*N
The corresponding number of training equations is
Ntrneq = Ntrn*O
Therefore, the condition for not over-fitting is
Ntrneq >= Nw = O + (I+O+1)*H
or equivalently,
H <= Hub = (Ntrneq-O)/(I+O+1)
where Hub is the lowest upper-bound.
If you need more that Hub hidden nodes there are two methods that mitigate over-fitting;
1. Validation stopping. The default data-division
yields
Nval = Ntst = round(0.15*N).
2. Bayesian regularization to prevent weights from
getting too large. This replaces the mse training
performance function with the regularization
combination
msereg = mse(error) + alpha* mse(weights)
alpha is automatically determined by the training algorithm.
The two ways to implement this are
a. net.trainFcn = trainbr
b. net.performFcn = msereg
Now, to answer your question. I don't know if you have overfit the net. However, valstop has prevented you from improving the training performance at the expense of nontraining performance.
Don't despair.
My approach is to design multiple nets in a double loop over number of hidden nodes (Hmin:dH:Hmax) with an inner loop over Ntrials random number states which dictate the random data-division and random initial weight assignments. Typically Ntrials = 10 and numH <= 10.
I have posted tens of examples. Searching
Hmin:dH:Hmax Ntrials
yields 70 hits.
Hope this helps.
Thank you for formally accepting my answer
Greg
