Reproduce “Learning Non-Convergent Non-Persistent Short-Run MCMC Toward Energy-Based Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Erik Nijkamp, Mitch Hill, Song-Chun Zhu, Ying Nian Wu
The authors introduce a short-run image generator that restarts from noise and applies a fixed number of Langevin steps guided by an energy model. Training matches parameter gradients on generated and observed images. On CelebA at 64×64, it reports FID 23.02, versus 32.71 for DCGAN and 79.09 for VAE. The comparison does not establish faster sampling or full-distribution matching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.