Use a backpropagation net to recognize the ten (hand drawn) digits. One approach would be to build a 4 x 6 array of points. When a digit is drawn on this grid it will cover some elements, giving them value 1, and miss others, value 0. This 24 element vector would be the input value for your net. You would build your own training vectors. Do the same task with a counterpropagation net; compare your results.