Reproduce “Distributional simplicity bias and effective convexity in 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
Aurélien Decelle, Alfonso de Jesús Navas Gómez, Beatriz Seoane
The authors analyze learning dynamics in energy-based models by rewriting binary-variable energies as effective fields and interactions of increasing order. Their theory predicts a conditional bias toward lower-order interactions, and restricted Boltzmann machine experiments report this progression, including transient pairwise terms in a synthetic system with only three-body interactions. The findings support the hierarchy under assumptions, but do not rule out spurious stationary points.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.