Accuracy estimation of neural networks by extreme value theory