Automatically training a Neural Network

[Niven]

I made a program that could recognize a 1 so I started work on a new program that would learn to recognize all the alphabet and the numbers 0-9. I'm trying to have to program automatically train itself (then I can save the weights of my neural network in a separate file so it doesn't have to retrain itself) but it's not going as well as I had hoped. It is able to recognize the different characters to some degree but gets confused with similarly shaped characters and characters displayed in different positions ( like if I display a 1 on the right side of the screen instead of the center ). I train it by going through and telling it that certain patterns correspond to a certain character. You can see my full source code here: http://www.mediafire.com/?ya0trt1qilwbin2 (look in inputscreen.cpp and net.cpp for training). What I want to know is what would be the best way for me to accurately train this network without having to manually display the different characters I want it to recognize and then tell it what I just drew?