Week 11: Generative models I: GANs

DATA 443/643 · Week 11

Module 4 — Generative models I: GANs

Notes

• 11.1 Generative Adversarial Nets(slides)
• 11.2 Training Dynamics(slides)
• 11.3 Mode Collapse and Evaluation(slides)

Labs

• Lab 11: A One-Dimensional GAN Sketch(PDF)

Videos

• Generative models: GAN setup

Additional readings

• Stanford CS231N 2025 L13: GANs give up on p(x) and sample in one pass
• Goodfellow et al., Generative adversarial nets
• Salimans et al., Improved techniques for training GANs

Discussion

• G(z)=+2 always. What is Bayes D(+2) and D(−2) on equal spikes at ±2?
• Why is a 99% discriminator often a dead generator?
• 92 fakes on A and 8 on B: which lab rule fires, and why can lossG still look small?
• Name one query a GAN cannot answer that an explicit likelihood model can.