Chapter 7. Generators in the Cloud

Creativity and the generation of new content were once the litmus test for true intelligence and what we thought consciousness may resemble. As it turns out, we were wrong. Creating “new” content isn’t that hard. In fact, we have created new content many times in previous chapters either as an effect or intent. Deep learning itself has made it possible to create new content across a wide variety of domains, from generating fake text from a chatbot to posting fake news on the web.

The explosion of deep learning has manifested many forms of content generation. Ironically, content generation systems are rarely the work of a single network and are often system trained in pairs or adversarially. We have already seen this with encoder/decoder architectures used for Seq2Seq learning. As we will see in this chapter, there are many other forms of adversarial learning to generate content.

It may seem that creating new content just for the purpose of creating new content would have a narrow appeal. Indeed, the ability to understand how content is created provides many insights into the domain problem itself, allowing us to tune new inputs and features in our models and ultimately how we understand data. As we will see, these types of insights have given us the ability to create things we’ve only dreamed of.

Generating content with AI is currently seen as the fastest-growing area of deep learning, an area we will focus on in this chapter. We will start by ...

Get Practical AI on the Google Cloud Platform now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.