ATML: Deep generative modelling
This track is an introduction to deep generative modelling algorithms, where we understand that term as follows:
- A generative model is a probability distribution, or generative process, that is derived from data so as to approximate the distribution that produced the data.
- A deep generative model is one that uses deep neural networks to represent components of the generative process.
- A deep generative modelling algorithm consists of: a choice of generative process, a family of distributions parametrised by neural networks to represent that process, and a learning algorithm to fit those networks' parameters to data.
This track will survey the main classes of deep generative modelling algorithms, particularly for high-dimensional data, and their applications.
Course information
Instructor: Esmeralda Whitammer; assistants: Sanghyeok Choi and Arran Carter. Thanks to Kirill Tamogashev and Rajit Rajpal for help preparing course materials for a previous edition of the course.
Lectures: Tuesdays 17:10-18:00, 1 George Square (Neuroscience), G.8 Gaddum Lecture Theatre.
Tutorials: Time, location, and staffing TBA, starting in Week 3. Theory exercises and code demonstrations will be posted the preceding week and linked to in the table below. Solutions to the theory exercises for weeks 2-5 will be made available in week 6 and for weeks 7-10 in week 11.
Recommended references: Probabilistic Machine Learning: An Introduction and Advanced Topics by Kevin P. Murphy. Chapter IV of the latter book concerns deep generative models.
Sample exam: see here.
Course schedule
Slides will be posted the evening before each lecture (possibly in draft form). In full pdfs, embedded videos work only in Adobe Reader and possibly some other programs. The compressed pdfs have images in lower quality and no embedded videos.
| Week | Date | Topic + recommended additional reading | Slides | Exercises |
|---|---|---|---|---|
| 1 | 12.01 | Introduction and overview | ||
| 2 | 19.01 | Preliminaries: probability review, distribution approximation and divergences | ||
| 3 | 26.01 | Latent variable models as generative models: autoencoders, variational autoencoders, hierarchical models, and their evaluation Kingma & Welling, “Auto-encoding variational Bayes” (ICLR, 2014) | ||
| 4 | 02.02 | Generative models with exact density evaluation: normalising flows, autoregressive models Papamakarios et al., “Normalizing flows for probabilistic modeling and inference” (JMLR, 2015) | ||
| 5 | 09.02 | Adversarial objectives for generative models: generative adversarial networks, mid-track summary Goodfellow et al., "Generative adversarial nets" (NeurIPS 2014) Arjovsky et al., "Wasserstein GANs" (ICML 2017) Huang et al., "The GAN is dead; long live the GAN!" (NeurIPS 2024) | ||
| Break | ||||
| 6 | 23.02 | Guest lecture by Zee Talat, “Evaluating generative AI” | ||
| 7 | 02.03 | Downstream uses of generative models: representation learning, conditioning and control; density-free evaluation of generative models Zeiler & Fergus, "Visualizing and understanding convolutional networks" (ECCV 2014) Theis et al., "A note on the evaluation of generative models" (ICLR 2016) Stein et al., "Exposing flaws of generative model evaluation metrics..." (NeurIPS 2023) | ||
| 8 | 09.03 | Diffusion and continuous-time models I: diffusion models as hierarchical VAEs | ||
| 9 | 16.03 | Diffusion and continuous-time models II: denoising and score matching, conditioning and guidance | ||
| 10 | 23.03 | Diffusion and continuous-time models III: continuous-time models | ||
| 11 | 30.03 | Revision or advanced topics depending on interest: more dynamics-based models, implicit (energy-based) models, discrete latent variable models and neurosymbolic methods, various applications, etc. | ||