Chapter 2: Deep learning
VAE vs GAN vs diffusion: quality, diversity, cost and FID
Interviewers ask you to compare VAEs, GANs (a new family) and diffusion models. They also ask how a GAN learns and fails, and how you would score generated samples. Here you learn the GAN game, the cost of one sample in each family, and FID, the standard score, with what FID cannot see.
What you will learn
- Explain the GAN minimax game and mode collapse
- Compare the three families on quality, diversity and sampling steps
- Compute FID and state what it misses
Key terms
- GAN
- a generative adversarial network: two networks that train against each other
- FID
- a score that compares generated samples with real samples; lower is closer
- mode collapse
- a generator that makes only a few values, so its samples lose spread
Prerequisites
Take Diffusion: noise forward, denoise backward first. It comes right before this lesson in the course.
Where this lesson picks up: You saw how a diffusion model removes noise step by step, with one network call per step. That left two questions open: how do other generators learn, and how can you tell which generator is better?
What comes next
Three models, three FID scores: which one do you ship, and what do you check before you trust the number?
Next lesson: Subword tokens and byte-pair encoding
More lessons in this part
- From autoencoders to VAEs: the ELBO and reparameterization 7 min
- Diffusion: noise forward, denoise backward 20 min
Take the lesson
Short cards, one idea at a time, with quick checks as you go.