Field Map

Energy-Based Models

Updated 2 days ago

A living map of energies, sampling, training, and compositionality in modern generative modeling.

Concepts

  1. Energy

    How models score configurations with unnormalized energy functions.

  2. Sampling

    MCMC, Langevin dynamics, and amortized ways to draw samples.

  3. Training

    Contrastive divergence, score matching, and modern objectives.

  4. Generation

    Using energies for images, sequences, and structured outputs.

  5. Composition

    Combining energies to compose concepts, constraints, and skills.