Build a working text generator from word counts, sculpt samples out of pure noise the way diffusion models do, then play the forger and detective game behind GANs.
3 units, about 3 hours. Free.
Build a text generator from bigram counts and explain why fluent output is not evidence of understanding.
2Run a toy diffusion sampler and explain forward noising, reverse denoising and coarse-to-fine generation.
3Play the adversarial game behind GANs and explain equilibrium, mode collapse and what it means for fake media.