Reproduce “On Uniform Convergence and Low-Norm Interpolation Learning” 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
Lijia Zhou, Danica J. Sutherland, Nathan Srebro
The authors analyze why low-norm predictors can generalize despite fitting noisy data exactly in Gaussian linear regression. They show that two-sided generalization-gap bounds over norm balls can fail, then study only interpolators whose Euclidean norms are near the minimum. For norms at most 1.1 times the minimum, expected worst-case population risk is at most 1.21σ² asymptotically; this is not a deep-network result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.