Implement an autoencoder backpropagation network and train it on the digits database. You can make one network and train it on all the digits (perhaps with an extra input telling it what the identity of the digit being encoded is, which goes everywhere). Or you can implement ten separate networks, each specialized for a particular digit.
Turn some representative examples of the reconstructed images (training and testing, expected digit and unexpected digit), some statistics (reconstruction error on-sample, off-sample, off-digit), and Hinton diagram(s) of your network(s).
I have placed the source code to my own backprop simulator on the web.
PS: Please please please work in teams!