In the paper, we show the results on MNIST and Fashion-MNIST datasets, we show the relationship between the distribution difference and the strength of errors on CIFAR-10 and SVHN datasets.