High-probability generalization bounds for pointwise uniformly stable algorithms

Abstract

Algorithmic stability is a fundamental concept in statistical learning theory to understand the generalization behavior of optimization algorithms. Existing high-probability bounds are developed for the generalization gap as measured by function values and require the algorithm to be uniformly stable. In this paper, we introduce a novel stability measure called pointwise uniform stability by considering the sensitivity of the algorithm with respect to the perturbation of each training example. We show this weaker pointwise uniform stability guarantees almost optimal bounds, and gives the first high-probability bound for the generalization gap as measured by gradients. Sharper bounds are given for strongly convex and smooth problems. We further apply our general result to derive improved generalization bounds for stochastic gradient descent. As a byproduct, we develop concentration inequalities for a summation of weakly-dependent vector-valued random variables.