Swarm GAN: Stabilizing Training of Generative Adversarial Networks via Swarm Intelligence

Abstract

Generative adversarial networks (GANs) have seen significant research interest over the past decade, yet core issues of training instability and mode collapse persist. This work proposes SwarmGAN, a novel GAN framework incorporating swarm intelligence to address these limitations. Specifically, swarm intelligence exhibits properties well-suited to enhance GAN training: emergent complex behaviors arising from simple individual agents, decentralized adaptability to instantaneous data and hyperparameters, and robustness through simple iterative interactions. SwarmGAN incorporates a particle swarm optimization algorithm to guide generator and discriminator updates. Convolutional neural network architectures and gradient penalties further ensure baseline generation quality and diversity. Extensive experiments over diverse image datasets demonstrate the effectiveness of SwarmGAN. Quantitative evaluations using Fréchet Inception Distance, Inception Score, Peak Signal-to-Noise Ratio, and Structural Similarity Index Score validate performance improvements across stability, sample quality, and convergence speed. The proposed integration of swarm intelligence into adversarial networks shows promising capability to address long-standing GAN challenges.