Generative Adversarial Networks Based On Maximum Mean Discrepancy

Generative Adversarial Networks(GANs)have now become one of the hottest research directions in the field of Artificial Intelligence.The basic idea of GAN originates from the Zero-Sum game in Game Theory.The generator and discriminator compete against each other.GAN has been successfully applied in the fields of Computer Vision,Speech,and Natural Language Processing.In addition,GAN has also played an important role in tasks such as Computer Vision and Games.However,at present,GAN is still in a stage of rapid development.There are problems such as unstable training,collapse of models,and inaccurate image generation.In this paper we proposes two new algorithms for the research of GAN:A generative adversarial network model based on moment matching of conditions(CMMD nets).This paper proposes a conditional generative adversarial network model based on Maximum Mean Discrepancy(MMD)to measure the difference in probability distribution of real samples and generated samples under the same class.The conditional probability density function cannot be obtained by directly using the parameter estimation method,In our method we instead seek the moment matching of the conditional distribution and extend the MMD to measure the class conditional distribution distance.At the same time,the MMD is directly applied to the feature space.Compared with GMMN,CMMD nets does not need large batch,which reduces the computation and includes the adversarial thinking of the original GAN.Compared to other MMD-based algorithms,CMMD nets do not require additional variational autoencoders(VAEs),Thus significantly reducing the computation and network complexity.Experimental results and analysis demonstrate that the proposed method can generate desirable samples.A generative adversarial network model based on joint distribution moment matching(JDMM GAN).Conditional GANs usually only minimize the margin distribution or conditional distribution of real data and generated data,but cannot guarantee that their margin distribution and conditional distribution can be reduced simultaneously.We propose a generative adversarial network based on joint distribution moment matching.It is difficult to estimate the probability density of data distribution,therefore we seek the moment matching of margin distribution and conditional distribution.We use MMD to measure the distance of probability distribution.JDMM GAN can minimize the distance between margin distribution and class conditional distribution at the same time.It can be used as a unified algorithm framework,and be applied to both unsupervised tasks and semi-supervised tasks.