Least Squares Generative Adversarial Networks

The paper targets the training instability of generative adversarial networks. The authors observe that a regular GAN, whose discriminator is a classifier trained with sigmoid cross-entropy loss, can suffer from vanishing gradients that hamper the generator. To address this, they propose Least Squares GANs, which replace the discriminator's loss with a least squares loss, and they show theoretically that minimizing the LSGAN objective is equivalent to minimizing the Pearson chi-square divergence.

LSGANs deliver two practical benefits over regular GANs: higher-quality generated images and more stable training dynamics. Experiments on the LSUN and CIFAR-10 datasets show that LSGAN-generated images are of better quality than those from regular GANs, and additional comparison experiments demonstrate the improved stability, making the least squares loss an influential and easy-to-adopt modification for GAN training.