Reproduce “Efficient Training of Energy-Based Models Using Jarzynski Equality” 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
Davide Carbone, Mengjian Hua, Simon Coste, Eric Vanden-Eijnden
The paper presents a weighted-walker method for training energy-based models with cross-entropy when sampling from the model distribution is difficult. Walkers follow unadjusted Langevin updates as the model changes, while changing weights correct their contribution; the ideal weighted expectation is exact under stated assumptions. In tests, it recovers separated-mode proportions and improves CIFAR-10 metrics over PCD, at similar training time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.