Meta-learning Enabled Score-Based Generative Adversarial Networks (GANs)

The capacity of Generative Adversarial Networks (GANs) to provide high-quality data has led to their significant attention. On the other hand, hyperparameter adjustment is a common part of GAN training, which may increase computing costs and result in less-than-ideal performance on certain tasks. Meta-Learning Enabled Score-Based GANs (MLS-GAN) is a new framework we provide in this study that combines meta-learning with score-based generative models. Our method employs meta-learning to enhance the training of score-based models via the collection of priors customized to individual tasks and tactics for dynamic adaptability. This improves the generative process’s generalizability and robustness, and it also allows for more efficient learning with fewer data and hyperparameter modifications. By conducting comprehensive tests on image synthesis and data creation tasks, we demonstrate that our Meta-SB-GANs are successful. The results reveal higher-quality samples, quicker convergence, and better transferability to other domains. Integrating meta-learning with generative models can achieve state-of-the-art performance with decreased computing resources, as shown by our findings.