Semester 4 · Computer vision
Deep Generative Models
https://epochschool.com/subjects/deep-generative-models
Learn how machines learn to create: models that capture the probability distribution of images and can sample new ones, score how likely an image is, or fill in what is missing.
You derive and build every major family, from autoregressive models, variational autoencoders and normalising flows to GANs, score matching, diffusion and flow matching, and then steer them with conditioning and classifier-free guidance. You learn to evaluate generators honestly with likelihood, Frechet distance and precision-recall, and to handle generated media responsibly with memorisation tests, watermarks and provenance. All five labs run on one CPU core with 8x8 digits and 2-D distributions, so every idea is implemented and tested from scratch. It builds on Deep Visual Recognition and Probability & Statistics and recalls the Semester 2 core instead of re-teaching it.
- Prerequisites
- Required: Finish this subject before you start. The subject builds directly on it.Recommended: Helpful background. You can start without it if you are willing to fill the gaps.
What you’ll be able to do
- LO1Formulate generative modelling as density estimation: relate maximum likelihood to KL divergence, compare forward and reverse KL, Jensen-Shannon, Wasserstein and Fisher divergences, and report likelihoods correctly in nats and bits per dimension with dequantisation.
- LO2Derive, implement and train autoregressive models (NADE, MADE, PixelCNN) with correct masking, exact likelihood evaluation and sequential sampling.
- LO3Derive the evidence lower bound for latent-variable models and apply it in EM and variational autoencoders, including the reparameterisation trick, importance-weighted bounds, posterior collapse and discrete latents.
- LO4Derive the change-of-variables formula and build normalising flows whose inverses and Jacobian log-determinants are exact and efficient.
- LO5Analyse and train adversarial models: derive the optimal discriminator and the divergence a GAN minimises, explain min-max dynamics and mode collapse, and apply stabilisers (non-saturating loss, Wasserstein critics with gradient penalty, spectral normalisation).
- LO6Derive score matching and denoising diffusion models (forward process, reverse posterior, ELBO, noise-prediction loss and its equivalence to denoising score matching) and implement their training and sampling.
- LO7Accelerate and steer sampling: derive DDIM and the probability-flow ODE, implement flow matching, conditional models and classifier-free guidance, and explain latent diffusion and diffusion Transformers with their compute trade-offs.
- LO8Evaluate generative models with likelihood, Frechet and kernel distances, precision and recall, and memorisation tests, and design an evaluation protocol whose conclusions are statistically sound.
- LO9Apply memorisation checks, watermarking and provenance to generated media, quantify detection error rates, and turn the consent, misuse and disclosure requirements taught in Responsible AI into engineering checks.
Chapters
The chapter and lab list for this subject is being finalised and will be published here.
Before you start
Required
Finish this subject before you start. The subject builds directly on it.
Probability & Statistics for ML
Recommended
Helpful background. You can start without it if you are willing to fill the gaps.
Image Formation and Classical Computer Vision
Where it leads
Subject 3 of 4 in the Computer vision track.
Previous: Deep Visual RecognitionNext: Vision-Language, 3D Vision and World Models
How it’s taught
You read short sections written for the page. Each ends with a check, and in-browser exercises let you practise. Graded labs run on your own cloud lab machine, CPU-friendly.
No videos. No live classes. No AI tutor.