As I understand it, the back-propogation of the weights adjustment involves comparing the result they got with the "correct" result from the data. The problem with data outside of training is that there's no "correct" result - nothing to judge against, so no method of back-propogating the weights corrections. But I could be wrong.