The risk tail of minimum-norm ReLU interpolation on the sphere
We study two-layer ReLU networks without biases that interpolate n independent Rademacher labels at independent uniform points of the sphere $\mathbb{S}^{d-1}$ with the smallest Euclidean norm of the weights. For univariate networks with biases Joshi, Vardi and Srebro showed that the expected squared risk of minimum-norm interpolants of noisy labels is infinite. In the bias-free model on the sphere, at every fixed sample size n ≥ 3, the tail of the squared risk is a power law with exponent (d-1)/2, from above and from below. So the expected risk is infinite in dimensions two and three and finite in dimension four and higher. The lower bound comes from a close pair of inputs with opposite labels and the other inputs on one side of the plane between them. The upper bound comes from an explicit interpolant whose cost is controlled by the smallest angle between two input lines.
Authors
- Yitzchak Shmalo (ORCID: https://orcid.org/0000-0002-2886-3049)
Institutions
- Hebrew University of Jerusalem (IL)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24258719
- Primary Topic
- Stochastic Gradient Optimization Techniques
- Type
- preprint