Reproduce “New Multiple Descent Phenomena Induced by Heterogeneous Data” 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
Mark-Anthony Moisescu-Pareja, Elvis Dohmatob
The authors show pooled data with different input covariances can produce source-dependent error peaks as model width increases. They study a random-feature layer with a trained linear readout. In synthetic-teacher experiments with 500 examples, an extra peak tracks source sizes of 125, 250, or 375, while the peak at width 500 stays fixed; these are squared prediction errors, not classification results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.