ParityGAN: class balance with hierarchical temporal memory

Abstract

Generative Adversarial Networks are efficient image generation models. They are still better performers in specific domains such as producing high resolution synthetic image data, medical imaging, image to image translation. GAN’s have some limitations such as, bias towards majority classes, lack of fairness constraint, mode collapse and class imbalance. To address these problems, We have introduced ParityGAN, a hybrid GAN architecture that deals with the existing limitations of GAN’s in class distribution and fairness. Our proposed model’s generator is the combined form of conditional GAN to ensure label aware generation, ResNet based DCGAN for high quality synthesis and spectral normalisation SNGAN for training stability. We have also introduced a novel approach that is TBM (temporal balanced memory) and updated version of it named ETBM (enhanced temporal balanced memory). ETBM monitors and adjusts class distribution through short, medium and long term memory. Our distributional fairness discriminator helps the generator to achieve fair class distribution and at the same time provide ETBM with class probabilities. We trained the model on an imbalanced CIFAR-10 dataset with a total of 45.9M parameters and achieved a FID score of 18.11. Our ParityGAN demonstrated that achieving fairness or reducing bias does not require sacrificing generation diversity.