Section 11 of 12 · 4 lessons

Generative Models

Recognising a picture and creating a new one are different problems. To generate examples, a model needs to learn something about how the training examples are put together. We will look at several ways to do this, from reconstructing compressed inputs to learning how to remove noise.

0 of 4 finished

Learning a Compact Representation

If a model compresses an input, what must it keep to reconstruct it? That question gives us a starting point for learning representations from which we might also generate new examples.

  1. 01AutoencodersIf we ask a model to compress an input and then reconstruct it, the reconstruction tells us what the compressed representation kept. An autoencoder learns that representation by trying to reduce what is lost.
  2. 02Variational AutoencodersA model that reconstructs known examples does not necessarily give us a good way to generate new ones. A variational autoencoder also shapes a distribution of compressed representations from which we can draw new codes.

Learning to Generate

A generator needs feedback about what to improve. We will compare learning from a model that distinguishes generated examples from real ones with learning to reverse a process that adds noise.

  1. 03Generative Adversarial NetworksA generator needs feedback about whether its examples resemble the training data. A GAN trains a second model to distinguish generated examples from real ones, giving the generator a changing source of feedback.
  2. 04Diffusion ModelsAdding noise to an example is easy to control. Learning to remove it gives us a useful training problem, and a trained reverse process can then turn a noisy starting point into a generated example.