Stochastic Nested Variance Reduction for Nonconvex Optimization

MaRDI QIDQ6303282arXivFDO

Authors Dongruo Zhou, Pan Xu, Quanquan Gu

Publication date 20 June 2018

Published in arXiv

Mathematics Subject Classification ID

Stochastic approximation (62L20) Stochastic programming (90C15) Nonconvex programming, global optimization (90C26)

arXiv classification

cs.LGmath.OCstat.ML

Abstract: We study finite-sum nonconvex optimization problems, where the objective function is an average of $n$ nonconvex functions. We propose a new stochastic gradient descent algorithm based on nested variance reduction. Compared with conventional stochastic variance reduced gradient (SVRG) algorithm that uses two reference points to construct a semi-stochastic gradient with diminishing variance in each iteration, our algorithm uses $K+1$ nested reference points to build a semi-stochastic gradient to further reduce its variance in each iteration. For smooth nonconvex functions, the proposed algorithm converges to an $epsilon$-approximate first-order stationary point (i.e., $|ablaF(mathbfx){|}_{2}leqepsilon$) within $ildeO(nlandepsilo{n}^{-2}+epsilo{n}^{-3}land{n}^{1/2}epsilo{n}^{-2})$ number of stochastic gradient evaluations. This improves the best known gradient complexity of SVRG $O(n+n^{2/3}epsilo{n}^{-2})$ and that of SCSG $O(nlandepsilo{n}^{-2}+epsilo{n}^{-10/3}land{n}^{2/3}epsilo{n}^{-2})$. For gradient dominated functions, our algorithm also achieves better gradient complexity than the state-of-the-art algorithms. Thorough experimental results on different nonconvex optimization problems back up our theory.

This page was built for publication: Stochastic Nested Variance Reduction for Nonconvex Optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6303282)