This code is using for active learning. one net using to predict the for data in active pool, the "probability"(used self entropy to present this) of wrongly predict by current model, another is just normal discriminte model. And I implement it to mnist data.