Research On Generative Adversarial Networks With Multiple Random Projections Based On Random Discard

With the rapid development of machine learning,especially deep learning and other related fields,the processing ability of computer for many difficult tasks has been greatly improved.However,in the process of implementing these algorithms,it is still very difficult to obtain a large number of labeled data,some of which need to be processed manually,some of which can not even be processed with human knowledge.Therefore,unsupervised learning has been paid more and more attention.In the field of deep learning,the most noticeable one is the Generation Adversarial Network(GAN).As a new unsupervised learning algorithm framework,researchers have paid more and more attention on Generation Adversarial Network because of its simple structure,wide application field,high quality of generated samples and no need to design loss function.Inspired by the theory of two person zero-sum game in game theory,Ian J.Goodfellow put forward this great structure of generation confrontation.Compared with other traditional machine algorithms,Generation Adversarial Network has better ability of feature expression and feature learning.At present,the application of Generation Adversarial Network in computer vision is particularly successful.However,there are still many defects and deficiencies in Generation Adversarial Network.Therefore,from the perspective of architecture,increasing the difficulty of discriminator identification as much as possible,this thesis proposes the architecture of training a generator against multiple discriminators at the same time,and discarding the discriminator randomly.In this architecture,the distribution is first projected into low dimensional space before discriminators' discriminations.Each discriminator focuses on different low-dimensional random projections,all of which contain all the features.By calculating the Wasserstein distance of each discriminator,the optimal discriminator is obtained,which ensures that the optimal discriminator is not discarded.At the same time,according to the performance of the discriminators,the probability of discarding is assigned to each discriminator.With this architecture,multiple discriminators can provide stable and learning value gradients to the generator.The generator learns according to the gradients,and then tries to generate samples consistent with the complete data distribution(approximate),so as to cheat all discriminators.The randomdiscarding of the discriminator can avoid the occurrence of over fitting and ensure the best quality of the whole discriminator group.In order to verify the quality of the image generated by the new architecture,experiments are carried out on three data sets MNIST(handwritten number),Cifar10(small object)and CelebA(celebrity face)in comparison with DCGAN.And evaluation is carried out from the perspectives of IS,FID and loss function value.The new architecture is better than DCGAN in quality and diversity.At the same time,in order to verify the impact of the number of discriminators on the architecture,conduct comparative experiments on the number of discriminators is 12,24,48 respectively.The quality of the generated image increases with the number of discriminators.