The accuracy of the system is considered to be how the trained network performs on the test set, ie, a set it hasn't seen during training.
The accuracy on the training set is only used during training, to update the weights and potentially determine when to exit training.