Publications · Preprint · arXiv 2023
Non-convex optimization via non-reversible stochastic gradient Langevin dynamics
arXiv preprint, 2023.
In brief
Takes the non-reversible Langevin idea to the stochastic-gradient setting: a non-reversible drift added to stochastic gradient Langevin dynamics leaves the target distribution unchanged while speeding up both sampling and non-convex optimization, and the paper quantifies the improvement in the convergence guarantees and the generalization bounds of the solutions found.
Cite
@misc{hu2023nonreversible,
title = {Non-convex optimization via non-reversible stochastic gradient Langevin dynamics},
author = {Yuanhan Hu and Xiaoyu Wang and Xuefeng Gao and Mert Gürbüzbalaban and Lingjiong Zhu},
year = {2023},
howpublished = {arXiv preprint arXiv:2004.02823},
}Relatedsame topics