Reproduce “How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning” 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
Max W. Shen, Mark Goldstein, Zichu Wang, Aahlad Puli, Rajesh Ranganath
The paper analyzes how stopgrad changes training updates and introduces a regression principle identifying stationary predictions with fixed points of a target operator. Applied to flow-map learning, it proves the true flow map is the unique stationary point for several objectives and derives idealized convergence. Slim ESD and LSD reduce peak memory, but slim LSD is worse at one-step CIFAR-10 sampling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.