Skip to main content

Generative AI

Variational autoencoders, GANs and diffusion models — the three families that learn to draw new examples that look like the training set, from handwritten digits to faces.

Course duration: 7h​

What you will learn​

  • Distinguish the discriminative approach from the generative one, and see why the second is a harder problem
  • Compress data with an autoencoder, then generate with a variational autoencoder
  • Train a DCGAN on MNIST and read its loss curves without fooling yourself
  • Diagnose mode collapse, apply the Wasserstein and spectral fixes
  • Understand the forward and reverse diffusion processes and sample from a small U-Net
  • Condition generation on a text prompt with CLIP embeddings and classifier-free guidance
  • Evaluate generation with FID and inception score, and know their limits
  • Think about copyright, provenance and consent when publishing a generative system

Prerequisites​

  • Deep learning fundamentals (course 07)
  • Convolutional neural networks (course 10)
  • Comfort with PyTorch and NumPy

Course modules​

  1. Discriminative and generative models
  2. Autoencoders and latent space
  3. Variational autoencoders
  4. GAN: generator versus discriminator
  5. Mode collapse and training stabilization
  6. Diffusion models: adding noise, then learning to denoise
  7. Text conditioning
  8. Quality metrics: FID and inception score
  9. Generating text, images and audio
  10. Copyright, provenance and responsible use

The thread through the course​

We start with handwritten digits — MNIST is small enough that a laptop trains a VAE, then a DCGAN, in a few minutes. Modules 2 to 5 stay with MNIST because it is the cheapest way to see the failure modes of each family. From module 6 we move to CelebA faces cropped to 64 by 64 pixels to train a small diffusion model, then module 8 evaluates that model with FID. Module 9 widens the picture to text and audio using existing models rather than training from scratch. Module 10 returns to the generated faces and asks who owns them, who can be identified in them and what the platform must disclose. Code is written in PyTorch throughout.

Assessment and certificate​

The course ends with a 40-question exam covering every module: reparameterisation trick, reading GAN loss curves, diagnosing mode collapse, the role of the diffusion timestep, guidance scale, when FID misleads, and the provenance obligations of a generative system. On success, your certificate of completion is issued immediately; its number can be verified by any third party on the platform.

Free courses, by contrast, end with a 5-question quiz and a certificate preview, without certification.