The Research Of Generative Algorithms Based On Adversarial

Last few years,Deep Learning had got great power in computer vision and natural language processing.Nowadays the generative adversarial nets and reinforcement learning had been thought of the most important working in deep learning.Different from discriminative models,learning generative models can help us understanding discriminative models and build more effective models.Because what we learned is a joint distribution,different from discriminative models,generate models always include more information.In this paper,we introduce the history of generative models and the development of generate adversarial nets.We also introduce some type of generative adversarial nets.Our main work mainly centered on the generative adversarial nets based on f-divergence.In generative models we use the difference between distributions in mathematic to analysis the models.Generative models is a hot topic in machine learning area,many generational models had been used in many applications.Some generative models such as RBM(Restricted Boltzmann Machine),VAE(Variational Auto Encoder).Because in research we found that use some noise in the train data can help us get more robust models,by some research we get the concept of adversarial.The DCGAN model,a variant of GAN help us build a good GAN struct,but in the theory of GAN,the development is slowly.Recently,some generative models based on Generative Adversarial Nets had beed used in computer version and natural language processing and had been proved good performance In the theory of GAN,especially the framework based on f-divergence make all variants of GAN together.In the application of GAN,the most useful is the GAN based on conditional.In this paper,we introduced a theory based on ALI,named f-ALI and analysis the theory of the f-divergence of the ALI.At last,we give some example of that.