- Learn
- Machine Learning
- Generative Models
ML-M14
Generative Models
Models that make new data: autoencoders and latent spaces, variational autoencoders, GANs and the diffusion models behind modern image generators.
Every model so far took an input and predicted a label or a number. Generative models learn what the data itself looks like, well enough to compress it, repair it, and produce new examples that were never in the training set.
You will start with autoencoders, which squeeze each input through a small code and rebuild it, and see what that code space looks like. Variational autoencoders make that space smooth enough to sample from and walk through. GANs train a generator against a critic and gave the first photorealistic faces. The module ends with diffusion models, which learn to remove noise one step at a time and now power most text-to-image systems.
Lessons
- ML-14.01 Autoencoders An autoencoder compresses input into a small latent code and reconstructs it, learning the most important structure of the data. Intermediate 12 min
- ML-14.02 Variational autoencoders A variational autoencoder learns a smooth latent space by encoding each input as a distribution and sampling from it, with a penalty that keeps the latent space well organized. Advanced 14 min
- ML-14.03 Generative adversarial networks A GAN trains two networks against each other: a generator that creates fake samples and a discriminator that tries to tell fakes from real data. Advanced 12 min
- ML-14.04 Diffusion models Diffusion models learn to reverse a gradual noising process. Advanced 16 min