Generative Models — Overview
Stub. The big picture of generative modeling: what a generative model is, the landscape of approaches (autoregressive, latent-variable, adversarial, score/diffusion-based), and how the rest of the course's topics fit into it.
The orienting lecture: a generative model learns (an approximation to) the data distribution itself, not just a decision boundary, which is what makes it possible to sample new, realistic data rather than only classify or predict. Surveys the major families — autoregressive models, latent-variable models (VAEs), adversarial models (GANs), and score/diffusion-based models — as different answers to the same underlying question (how do you represent and sample from an intractable high-dimensional distribution), setting up the course’s actual trajectory: VAEs, then diffusion models as a hierarchical generalization of VAEs, then their continuous-time (SDE/flow) formulation and applications.
Outline (TODO — flesh out each)
- Course logistics and big picture
- What is a generative model? Density estimation vs. sampling
- Overview of generative model families — autoregressive, latent-variable, adversarial, score-based/diffusion