Improving GAN Training via Binarized Representation Entropy (BRE) Regularization