Reproduce “How to scale your HEP ML models: A recipe for robust architecture comparisons at scale” 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
Matthias Vigl, Nikita Pond, Jackson Barr, Alexander Froch, Dan Guest, Nicole Hartman, Michael Kagan, Lukas Heinrich
The authors develop a procedure for tuning learning rates and batch sizes as training horizons change, then use it to compare compute-optimal model and dataset allocations for simulated jet-flavor taggers. Three fitting methods find near-equal allocation exponents: model size scales as C^0.52±0.01, and dataset size as C^0.48±0.01. This power-law frontier appears only after substantial data in the studied setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.