In one paragraph
Tomczak builds the main families of deep generative models (autoregressive models, flows, latent-variable models and VAEs, energy-based and diffusion models) from the same starting point: a model of $p(x)$ that we can learn and sample from. Each chapter pairs the derivation with short PyTorch code.
My review
[Your verdict, once you finish the book.]
What I learned
- Classification models $p(y \mid x)$; generative modeling models the data itself, through $p(x, y) = p(y \mid x)\,p(x)$.
- A confident classifier is not a certain one. Without $p(x)$, it cannot tell when an input is unlike anything it was trained on.
- [More as you read.]
Case study: how I used it
I work through the book in marimo notebooks, writing every line of code myself. Each section becomes a numbered problem. For example, Problem 1.1 asks when a confident classifier is actually certain, and the notebook answers it with a small experiment. [Link the repo, and where the book's ideas show up in your research.]
Going deeper: references worth following
- D. P. Kingma and M. Welling. Auto-encoding variational Bayes. ICLR, 2014.
- G. Papamakarios et al. Normalizing flows for probabilistic modeling and inference. JMLR 22, 2021.
- K. P. Murphy. Probabilistic Machine Learning: Advanced Topics. MIT Press, 2023.