In this assignment, you are to find a way to effectively reduces the size of your data.
Take the farsi hand-written digits and devise a way to reduce the size of the data such that you still can classify the samples with and accuracy above 97%.
What's the smallest number of features, with wich you still can keep the accuracy above 97%? (Just try some networks and give a rough answer.)
Try to combine your idea with convolution and see how that affects the result.