In this chapter, we looked at generative adversarial networks, or GANs, as a way to build DNNs that can generate unique content based on copying or extracting features from other content. This also allowed us to explore unsupervised training, a method of training that requires no previous data classification or labeling. In the previous chapter, we used supervised training. We started with looking at the many variations of GANs currently making an impression in the DL community. Then we coded up a deep convolutional GAN in Keras, followed by the state-of-the-art Wasserstein GAN. From there, we looked at how to generate game textures or height maps using sample images. We finished the chapter off by looking at two music-generating GANs that can generate original MIDI music from sampled music.

For the final sample, we looked at music generation with GANs that relied heavily on RNNs (LSTM). We will continue our exploration of RNNs when we look at how to build DL chatbots for games.