nnstart Neural Net toolbox and validation ROC

Hi,

I have been training neural networks for classification with nnstart. I get perfect training restults 0% false positives (FP) and 100% true positives (TP). The testing set performs quite less well but still acceptable, usually i can get up to 60% TP for 40% FP (and sometimes 80% TP). However, the Validation set ROC is usually very bad: i.e., random or worse. Can someone help me understand what it means ? What does it mean for the validation ROC to be bad and for the training ROC to be perfect ?

P.S. I use nn classification with usually about 100-200 samples split by default with 70% training set, 15% validation, 15% test. And default params, scaled conjugate gradient back-propagation & cross entropy minimization, 1000 hidden layers with sigmoid activation and last layer with softmax squashing function.