Reproduce “Distributional Generalization: A New Kind of Generalization” 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
Preetum Nakkiran, Yamini Bansal
The paper studies distributional generalization: whether a classifier’s train and test outputs have similar distributions, beyond matching average error. In a CIFAR-10 experiment, a WideResNet-28-10 trained to zero error on Animal/Object labels, with 30% of cat labels flipped, reproduces roughly the same localized errors on test cats. This behavior is not Bayes-optimal and is not established for all models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.