Reproduce “A Solvable Model of Adaptive Learning Rate Rescaling: Acceleration, Stability & Scaling” 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
Itay Lavie, Clarissa Lauditi, Cengiz Pehlevan
The authors develop DMFT for normalized stochastic-gradient descent, which fixes update length while retaining gradient direction. In the infinite-width, infinite-batch continuous-time limit, normalization changes the loss exponent from r to 2r/(1-r) when r<1, accelerating early learning across spectral modes. At finite step size, the effective learning rate reaches a stability ceiling. Batch or width may reduce steps without guaranteeing compute savings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.