Design and Visualization of Guided GAN on MNIST dataset

Abstract

In this paper, we propose a hybrid model aiming to map input noise vector to the label of the generated image by Generative Adversarial Network (GAN). This model mainly consists of a pre-trained Deep Convolution Generative Adversarial Network (DCGAN) and a classifier. By using the model, we visualize the distribution of two-dimensional input noise leading to specific type of generated image after each training epoch of GAN. The visualization reveals the distribution feature of the input noise vectors and the performance of the generator. With this feature, we try to build a Guided Generator which can produce a fake image which we exactly need. Two methods are proposed to build the Guided Generator. One is the most significant noise (MSN) method, and another is utilizing labeled noise. MSN method can generate images precisely but with less variations. In contrast, labeled noise method can have more variations but slightly less stable.