Think it would have been more interesting to see you use the Cifar-10 dataset instead as it has the same number of classes, image size, test and training size, while also being harder and more realistic. In this way, we could really see the power of the method you are using. MNIST can be solved to 99% without a CNN.