Reproduce “Transfer Learning in Infinite Width Feature Learning 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

Clarissa Lauditi, Blake Bordelon, Cengiz Pehlevan

The authors develop a transfer-learning theory tracking how pretraining and target training reshape internal features through adaptive kernels. In an isotropic linear model, population-data pretraining improves target test loss over a random-start probe under specified feature-learning and task-alignment conditions. Finite source samples can cause negative transfer. A small image-regression experiment also finds lower held-out target loss in its tested setting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.