FourierSpectrumDiscrepanciesinDeepNetwork GeneratedImages

For the majority of the datasets, 10% of the images were used for training while the remaining 90%wereusedfortesting tohighlight therelativelylownumber oftraining examplesrequired for classification.