Single-loop Algorithms for Stochastic Non-Convex Optimization with Weakly-Convex Constraints

Ming Yang [email protected] · 2025

Read Single-loop Algorithms for Stochastic Non-Convex Optimization with Weakly-Convex Constraints by Ming Yang [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-15243.txt.

Description

Ming Yang and co-authors present single-loop algorithms for stochastic optimization problems whose objective and constraints are weakly convex, published in 2025. Problems with several functional inequality constraints matter widely in machine learning. Existing methods for this weakly convex class often converge slowly, or need a double-loop design that is costly to run.

Their method builds on the classical exact penalty approach, using a hinge-based penalty. That choice permits a constant penalty parameter, a simplification that earlier methods lacked. The team proves a state-of-the-art complexity bound for finding an approximate solution of the optimality conditions. They extend the method to finite-sum compositional objectives, common in AI applications, with improved complexity there as well. Tests cover fair learning and continual learning tasks.

Catalog record

CREATORMing Yang [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE110 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-15243

Browse the full library · About the Tower of Babel