Latent Dirichlet Allocation in Generative Adversarial Networks

Mode collapse is one of the key challenges in the training of Generative Adversarial Networks(GANs). Previous approaches have tried to address this challenge either by changing the loss of GANs, or by modifying optimization strategies. We argue that it is more desirable if we can find the underlying structure of real data and build a structured generative model to further get rid of mode collapse. To this end, we propose Latent Dirichlet Allocation based Generative Adversarial Networks (LDAGAN), which have high capacity of modeling complex image data. Moreover, we optimize our model by combing variational expectation-maximization (EM) algorithm with adversarial learning. Stochastic optimization strategy ensures the training process of LDAGAN is not time consuming. Experimental results demonstrate our method outperforms the existing standard CNN based GANs on the task of image generation.