- Who is the adversary in a GAN?
The generator. It acts as a model whose sole purpose is to make the critic fail; it is the critic's adversary.
- Why is the generator model bigger than the critic?
This is not always the case. The models discussed here were more interesting as generators of data. However, we could use the critic and retrain it for classification, in which case, the critic model might be bigger.
- What is adversarial robustness?
It is a new field in deep learning tasked with researching ways to certify that deep learning models are robust against adversarial attacks.
- Which is better – a GAN or a VAE?
This depends on the application. GANs tend to produce more "interesting" results than VAEs, but VAEs are more stable. Also, it is often faster to train a GAN than a VAE.
- Are there any risks associated with GANs?
Yes. There is a known problem called mode collapse, which refers to the inability of a GAN to produce novel, different...
