Reproduce “The Multiscale Single-Index Model: An MSE Analysis of Online SGD Dynamics” 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
Gordon Dai, Joan Bruna
The authors analyze joint online spherical SGD with mean-squared error (MSE) in a two-hidden-layer, multiscale model, where both learned directions start with small overlaps with planted features. Under non-degeneracy, delocalization, and favorable initialization, they prove weak recovery in O(d log² d) samples, followed by strong recovery in O(d log(1/ε)) additional samples. The result is limited to K=2; deeper settings remain open.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.