English

Generalization error in high-dimensional perceptrons: Approaching Bayes error with convex optimization

Machine Learning 2021-02-18 v2 Disordered Systems and Neural Networks Machine Learning Statistics Theory Statistics Theory

Abstract

We consider a commonly studied supervised classification of a synthetic dataset whose labels are generated by feeding a one-layer neural network with random iid inputs. We study the generalization performances of standard classifiers in the high-dimensional regime where α=n/d\alpha=n/d is kept finite in the limit of a high dimension dd and number of samples nn. Our contribution is three-fold: First, we prove a formula for the generalization error achieved by 2\ell_2 regularized classifiers that minimize a convex loss. This formula was first obtained by the heuristic replica method of statistical physics. Secondly, focussing on commonly used loss functions and optimizing the 2\ell_2 regularization strength, we observe that while ridge regression performance is poor, logistic and hinge regression are surprisingly able to approach the Bayes-optimal generalization error extremely closely. As α\alpha \to \infty they lead to Bayes-optimal rates, a fact that does not follow from predictions of margin-based generalization error bounds. Third, we design an optimal loss and regularizer that provably leads to Bayes-optimal generalization error.

Keywords

Cite

@article{arxiv.2006.06560,
  title  = {Generalization error in high-dimensional perceptrons: Approaching Bayes error with convex optimization},
  author = {Benjamin Aubin and Florent Krzakala and Yue M. Lu and Lenka Zdeborová},
  journal= {arXiv preprint arXiv:2006.06560},
  year   = {2021}
}

Comments

11 pages + 45 pages Supplementary Material / 5 figures, v2 revised and accepted at NeurIPS

R2 v1 2026-06-23T16:14:37.603Z