Brainstorming Generative Adversarial Networks (BGANs): Towards Multi-Agent Generative Models with Distributed Private Datasets

To achieve a high learning accuracy, generative adversarial networks (GANs)\nmust be fed by large datasets that adequately represent the data space.\nHowever, in many scenarios, the available datasets may be limited and\ndistributed across multiple agents, each of which is seeking to learn the\ndistribution of the data on its own. In such scenarios, the agents often do not\nwish to share their local data as it can cause communication overhead for large\ndatasets. In this paper, to address this multi-agent GAN problem, a novel\nbrainstorming GAN (BGAN) architecture is proposed using which multiple agents\ncan generate real-like data samples while operating in a fully distributed\nmanner. BGAN allows the agents to gain information from other agents without\nsharing their real datasets but by ``brainstorming'' via the sharing of their\ngenerated data samples. In contrast to existing distributed GAN solutions, the\nproposed BGAN architecture is designed to be fully distributed, and it does not\nneed any centralized controller. Moreover, BGANs are shown to be scalable and\nnot dependent on the hyperparameters of the agents' deep neural networks (DNNs)\nthus enabling the agents to have different DNN architectures. Theoretically,\nthe interactions between BGAN agents are analyzed as a game whose unique Nash\nequilibrium is derived. Experimental results show that BGAN can generate\nreal-like data samples with higher quality and lower Jensen-Shannon divergence\n(JSD) and Fr\\`echet Inception distance (FID) compared to other distributed GAN\narchitectures.\n