Stanford CS236 Deep Generative Models: Lecture 1

A/Prof Stefano Ermon @Stanford gave this fantastic course on DGMs in the Spring of 2024. I thought about writing some notes in Apple Notes, then moved them here. All credit goes to Prof Ermon, but sometimes I may add some clarifications.

The Fundamentals

Deep Generative Models (DGMs) tackle a core challenge in AI: making sense of complex, unstructured inputs. This applies to fields like Computer Vision, NLP, Computational Speech, and Robotics.

Key Concept

As Richard Feynman said, "What I cannot create, I do not understand." In DGMs, we flip this:

"What I understand, I can create."

Model Structure

Most models follow a similar pattern: generation + inference.

Statistical Generative Models

These are probability distributions p(x) learned from data. They incorporate:

• Data samples
• Prior knowledge (parametric form, loss function, optimization algorithm)

In essence, we're creating a simulator for the data-generating process:

Control signals → data simulator → new data points
Potential data points → data simulator → Probability values

data simulator ⇔ statistical model ⇔ generative model

This approach allows us to build models that can both generate new data and infer probabilities, opening up new possibilities in AI and machine learning.