Classification-accuracy monitored backpropagation

For IC diagnostic purposes, the classification accuracy of training patterns (CAT) and the classification accuracy of unseen patterns (CAU) can be used as a measure of feedforward neural network (FFN) performance. To maximize FFN generalization performance, a CAU monitored back-propagation (BP) training technique is investigated and compared with the conventional minimum mean squared error criterion. To prevent over-training, an extra set of untrained data is used to monitor the generalization accuracy of the FFN. Using this technique, the trained FFN is optimized for generalization. The experiment has shown that the CAU monitored BP training algorithm improved the FFN classifier generalization accuracy compared to the minimum mean squared error criterion