The plot shows that the training error, that is the error evaluated on the points belonging to the training set, keeps on decreasing as the number of iterations of the training algorithm increases. But the generalisation error, that is the error computed on a number of solutions that don't belong to the training set (generalisation set) start increasing after a certain number of iterations. The generalisation set can be used to perform the so called cross validation process. The network keeps lowering the training error, the accuracy is very good, but when the error computed on the generalisation set start raising, we have to stop our training procedure (avoiding over-training).

---