Toy model of a Generative Adversarial Network (see: https://arxiv.org/abs/1406.2661) where:
- the generator is trained to generate fake normal distributions from noise (uniform distributions).
- the discriminator is trained to detect if the given normal distributions are real or fake/generated.
The uniform distributions (noise) are generated on-the-fly with a uniform distribution, which is quite a common process. The normal distributions -- the training data -- are also generated on-the-fly, which has 2 advantages:
- this simulate infinite data
- this makes the code much shorter -- no need to read, format, parse external data -- so we can focus more on the GAN itself!