Because I was experimenting with a neural network I thought it would be cool to have an even larger test- and training set. So I quickly hacked together this Python tool that creates labeled 28×28 images my neural network could understand.

It uses several fonts to create the little digit images and uses random rotation and translation. It can also add some noise to the pictures.

I was impressed that my neural network could decipher those digits with ease as well (be it with a lower accuracy of about 94% though) !