Reproduce “Fractional Underdamped Langevin Dynamics: Retargeting SGD with Momentum under Heavy-Tailed Gradient Noise” 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

Umut Şimşekli, Lingjiong Zhu, Yee Whye Teh, Mert Gürbüzbalaban

Fractional underdamped Langevin dynamics (FULD) retargets momentum updates under symmetric α-stable gradient noise by transforming velocity before it moves parameters, reducing the influence of large velocities. The authors prove its continuous-time position marginal has the intended Gibbs distribution. In a quartic-potential Cauchy-noise simulation, it tracks the target better than uncorrected dynamics; neural-network results are mixed and assume α-stable noise.

Summary by alphaXiv.

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