Section 3a: f-divergences
One convex function f turns a density ratio into a divergence. Forward KL, reverse KL, Jensen–Shannon, and total variation are four choices of f.
Writing
Generative AI
Notes on AI platforms, agent quality, and how software engineering is changing.
Section 2: The general principle of generative models
Pick a parametric family, define a divergence, and minimise it using only samples.
Section 1: The dataset and the problem
The hidden distribution, the dataset I actually get, and why generative modeling starts there.