Machine Learning with Generative Adverserial Network

Generative Adversarial Networks have been presented to train generative models as they have pose a certain level of exertion and difficulty in the past years. Comparatively, GAN has received less attention. This model shows that it can generate MNIST (Modified National Institute of Standards and Technology database) digits conditioned on class labels. Translating an image from a certain source image or the original image to a clearer and higher resolution target domain is the approach introduced in this paper. This work presents a non-conventional way to resolve the problem presented in digitizing handwritten digits. This approach is supposed to be easier to implement and understand while having a high efficiency and accuracy as compared to the existing models. It is shown that MNIST digits conditioned on class labels can be generated by this model. Both the generator and discriminator are to be conditioned. The idea is to nurture both the generator and discriminator together and gradually. Beginning from a low-resolution image, more layers are added so that the model can increase the definite finer details and progress towards a clearer image as the training progresses.