Learnable GAN Regularization for Improving Training Stability in Limited Data Paradigm

Generative adversarial networks (GAN) are generative models that require large amounts of training data to ensure a stable learning trajectory during the training phase. In the absence of sufficient data, GAN suffers from unstable training dynamics that adversely affect the quality of generated data. This behavior is attributed to the adversarial learning process and the classifier-like functioning of the discriminator. In data-deficient cases, the adversarial learning procedure leads to the discriminator memorizing the data instead of generalizing. Due to their wide applicability in several generative tasks, improving the GAN performance in the limited data paradigm will further advance their usage in data-scarce fields. Therefore to circumvent this issue, we propose a loss-regularized GAN, which improves the performance by forcing a strong regularization on the discriminator. We conduct several experiments using limited data from the CIFAR-10 and CIFAR-100 datasets to investigate the effectiveness of the proposed model in overcoming discriminator overfitting in the lack of abundant data. We observe consistent performance improvement across all the experiments compared to state-of-the-art models.