Reproduce “Silent Alignment in Two-Layer Linear Networks: A High-Dimensional Perspective” 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
Simon Martin, Francis Bach, Giulio Biroli
The authors analyze gradient flow in a whitened two-layer linear network with small isotropic imbalance. They track predictor singular directions separately from singular values, showing that directions can begin aligning with the input–output cross-moment matrix while loss remains near its initial value. Amplitude learning follows later; this empirical-fit result does not establish a generalization benefit.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.