Decomposition of Random Vector of GAN and Basis Vector Assembly for Chinese Character Generation

Abstract

Ian et al published the paper generative adversarial networks (GAN) [1], which consists of two networks one is generator, the other is discriminator. Its principle is that z-vector with a length of 100 then each element along the Gaussian distribution, is inputted into generative network than it makes fake image, which is similar real image what discriminator wouldn't distinguish. In adversarial relation, they are elaborated gradually with training. Finally, discriminator can't distinguish them. In other words, if they have a sufficient training time, z-vector will follow real data set distribution almost exactly. After that, Alec et al's deep convolutional generative adversarial networks (DCGAN) [2] introduced the Vector-Arithmetic, which uses the feature of sufficiently trained GAN's z-vector. It showed that each z-vectors has extraction of image's feature and has relation. However, it revealed that z-vector reflect approximate image characteristics but was not analyzed in practice what z-vector's elements has relation with them. In this paper, we analyzed the relation of z-vector's elements and introduce the Basis Vector Decomposition & Assembly (BVD & BVA).