| As one of the important achievements of unsupervised deep learning,generative adversarial networks have attracted more and more scholars’ attention and research.Traditional generative adversarial networks adopt the idea of binary zero-sum games,which generate convincing images by constructing a generator and a discriminator,and establishing adversarial relationships.The core of generative adversarial networks is the measurement of distribution and the construction of network model structures.This paper first proposes a new measurement method-view distance,and then constructs a new generative adversarial network structure-multiple generative adversarial network.The main contributions of this article are as follows:1)Considering that in the high-dimensional data space,it is difficult to describe the similarity or distance between two data points by Euclidean distance to describe the internal structure of the data.First,the rationality of the view distance and the related properties of the view distance matrix are proved in theory,the concept of similarity differentiation gain is proposed,and then the view distance is applied to K-Means algorithm.On classic manifold learning datasets,the K-Means algorithm based on view distance has achieved relatively good results,not only being able to cluster data according to its own structure,but also having a neat dividing line between categories.This article also tested the classification accuracy and clustering effect of the K-Means algorithm based on view distance on real-world datasets.The experimental results show that on most datasets,the K-Means algorithm based on view distance has significantly improved its classification accuracy and clustering effect compared to the original K-Means algorithm,indicating that view distance has certain theoretical and application advantages.2)Inspired by the idea that view distance measures the similarity between data points from multiple dimensions,we further propose a generative adversarial network model.By adding a generator,not only do we establish adversarial relationships between the discriminator and the new generator,but also establish adversarial and collaborative relationships between the two generators,transforming the binary minimax game of the original generative adversarial network into a multivariate adversarial collaborative relationship.3)It is proved theoretically that the global optimality of the solution and the two generators obtained from the generated antagonism network,one of which has better performance than the inference of the traditional binary generated antagonism network and the convergence of the algorithm.Finally,the effect of the new model is tested on the MNIST,CIFAR10 and Celeb A data sets.The experimental results show that the loss function of the optimal generator and the generated image quality(FID value)of the multiple generation adversarial network are smaller than the traditional binary generation adversarial,which shows the effectiveness of the model. |
