Similarity-Based Gossip Learning for Generative Adversarial Networks

Similarity-Based Gossip Learning for Generative Adversarial Networks

연구 분야: Artificial Intelligence

학회: 2023 15th International Conference on Communication Software and Networks (ICCSN)

Generative Adversarial Networks (GANs) have shown remarkable results in tasks such as image generation and data augmentation, but traditional centralized training methods often require a large number of computational resources and high-speed network connections, making it difficult to apply them to large-scale and distributed scenarios. In this paper, we propose a distributed GAN training method based on Gossip Learning, which realizes decentralized communication and collaborative learning among clients. We also use the Meta-Learning framework to improve the model's generalization ability in few-shot settings. We employ similarity-based neighbor selection in the Gossip algorithm to make collaborative learning more efficient. Experiments on various datasets showed that compared to traditional centralized training methods, using Gossip Learning significantly improves the training performance of GAN models and can be scaled to large-scale distributed training.