Generating diverse synthetic datasets
In this section, you will learn different methods of generating diverse synthetic datasets. We will discuss the following:
- Latent space variations
- Ensemble synthetic data generation
- Diversity regularization
- Incorporating external knowledge
- Progressive training
- Procedural content generation with game engines
Latent space variations
Latent space usually refers to a high-dimensional space where the training data is represented in a more abstract or compact way. Deep learning with many layers is designed to make the features in the latent space capture more semantic and conceptual information. For more details, please refer to Chapter 1. Thus, these features, in that space, convey encoded information about the problem through the ML model during the training stage. We may not be able to directly link the changes in the latent space to the changes that will happen on the generated images in models such as GANs....
