Energy-Based Models: Lower Energy, Higher Probability
An interactive live film: 3D double-well energy surface where valleys are probability peaks; the score vector field is the negative energy gradient, acting as a physical force field pulling samples into high-probability modes.
Narration
An energy-based model defines a probability distribution through an unconstrained scalar energy function E of x. Unlike standard densities, it requires no tractability constraints on the network.
Look at this three-dimensional energy surface. High plateaus represent unlikely, low-probability configurations. Deep bowls and valleys mark stable, favored states.
The rest of the narration is in the book's Companion.
