Fig. 1. Protocol for trials in GAN performance experiment. The GAN learns L from a training sample S. The GAN error d(L, T) is then estimated by comparing samples Q and R from L and T, respectively.