Reproduce “M-EBM: Towards Understanding the Manifolds of Energy-Based Models” 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
Xiulong Yang, Shihao Ji
M-EBM initializes repeated image sampling from a single Gaussian estimated over the training set, rather than unrelated noise. It combines this with changes to M-JEM’s training procedure and energy head. On CIFAR10, M-JEM with five steps reports FID 12.1, versus 38.4 for JEM with twenty steps. This comparison does not isolate initialization, and stability is not guaranteed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.