Reproduce “Investigating the Early Signatures of Feature Learning in Deep Linear Networks” 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
Shaunak Bhandarkar, Boris Hanin
The authors derive how one gradient step changes feature kernels in an infinite-width deep linear network. The step adds a target-dependent rank-one term whose strength grows with depth. They analyze kernel ridge regression in a spiked Gaussian setting and give a sufficient signal-to-noise window for beating every optimally tuned fixed-feature baseline; the result does not establish gains outside that regime.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.