Lesson 313
Generative Models
VAEs · GANs · Diffusion
1:00How VAEs, GANs, and diffusion models learn a data distribution and sample brand-new outputs — and the trade-offs of fidelity, diversity, and speed each family makes.
By the end, you can
- Explain the difference between discriminative and generative models using P(y|x) vs P(x).
- Describe what a latent space is and why smooth interpolation is a sign of good organization.
- Trace the VAE pipeline: encoder to distribution, reparameterization trick, decoder, KL term.
- Explain why the reparameterization trick is necessary for gradient-based training.
- Describe the GAN minimax game, the Nash equilibrium endpoint, and why training is unstable.
- Define mode collapse and explain why it limits GAN diversity.
- Explain the forward noising and learned reverse denoising processes in diffusion models.
- Compare VAEs, GANs, and diffusion models on fidelity, diversity, and sampling speed.
- Explain why FID and Inception Score are imperfect and can be gamed.
- Distinguish between "sampling a distribution" and "copying training data," and describe when memorization is a real risk.
Up next in AI, Machine Learning & Course Review
