Training of Generative Adversarial Networks using Particle Swarm Optimization Algorithm

In this paper, a particle swarm optimization (PSO) based solution is proposed for the training of generative adversarial networks (GANs). Conventional GAN networks take around 5x times more number of iterations to generate plausible images compared to the proposed method, thereby increasing the simulation time and decreasing the Frechet Inception Distance (FID) score. To overcome the problems of non-convergence and mode collapse associated with the conventional GANs, proposed work uses a PSO algorithm to stabilize the inertia weights during the training duration followed by conventional optimization method for the remaining iterations. The proposed solution is implemented on Nvidia Tesla VI00-PCIE-16GB GPU, using tensorflow and keras. The efficiency of the proposed solution is verified using MNIST dataset. The results showed that the iteration at which images are generated for the proposed method is faster as compared to the conventional GAN architectures, quantified with lower FID score.