Generative Adversarial Network

Generative Adversarial Network (GAN)

GAN is quite different from predictive deep learning models that commonly produce predictions, such as tomorrow’s weather or stock market prices. Instead, the objective of GAN is to learn a sample space, and then produce synthetic samples from the learning.

Figure 1-1. Generated handwritten digits by GAN.

The most general application of GAN is producing synthetic images, like the produced numbers in Figure 1-1. GAN is generally applied to image datasets so that GAN can produce synthetic images. However, There is a crucial limitation in the ordinary GAN in which we cannot produce a desired sample, such as a specific number.

Controllable GAN

Figure 1-2. Generated face images by ControlGAN.

In order to handle the limitation of conditional generation, we proposed Controllable GAN (ControlGAN). ControlGAN can produce desired samples that correspond to input values of the conditions, as shown in Figure 1-2.