Reproduce “Investigating the generative dynamics of energy-based neural networks” 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

Lorenzo Tausani, Alberto Testolin, Marco Zorzi

The paper tests whether changing an RBM’s hidden-state initialization makes top-down generation visit more digit classes. It compares single-digit biasing with intersection and double-label chimera constructions in 100-step MNIST trajectories. Intersection averaged 2.951 ± 0.099 visited states versus 1.779 ± 0.211 for baseline; double-label reached 2.104 ± 0.122 but was not significant. No trajectory explored all classes, and non-digit labels persisted.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.