Reproduce “Generalized Energy Based Models” on OpenResearch

About this paper

Michael Arbel, Liang Zhou, Arthur Gretton

GEBMs combine a generator that learns data support with an energy function that reweights probability mass on that support. During sampling, MCMC adjusts the generator’s latent inputs according to the energy before producing images. On CIFAR-10 with an SNGAN ConvNet, GEBM-ULA achieved FID 23.02 versus 32.03 for KALE-GAN under the reported 1,000-iteration setup, but reweighting cannot create missing support.

Key insights

• A unified framework that combines the strengths of implicit models (for manifold learning) and explicit models (for density refinement) results in more effective generative models.
• The proposed KL Approximate Lower-bound Estimate (KALE) provides a theoretically sound, Lipschitz, and differentiable objective for training GEBMs, ensuring stable and effective gradient-based optimization.
• Efficient and theoretically guaranteed sampling is achieved by performing MCMC in the low-dimensional latent space of the base distribution, with convergence rates scaling favorably with the latent dimension (O(exp(−dim(Z) * t))).

Summary by alphaXiv.