We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

math

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Stochastic Second-Order Methods Provably Beat SGD For Gradient-Dominated Functions

Abstract: We study the performance of Stochastic Cubic Regularized Newton (SCRN) on a class of functions satisfying gradient dominance property which holds in a wide range of applications in machine learning and signal processing. This condition ensures that any first-order stationary point is a global optimum. We prove that SCRN improves the best-known sample complexity of stochastic gradient descent in achieving $\epsilon$-global optimum by a factor of $\mathcal{O}(\epsilon^{-1/2})$. Even under a weak version of gradient dominance property, which is applicable to policy-based reinforcement learning (RL), SCRN achieves the same improvement over stochastic policy gradient methods. Additionally, we show that the sample complexity of SCRN can be improved by a factor of ${\mathcal{O}}(\epsilon^{-1/2})$ using a variance reduction method with time-varying batch sizes. Experimental results in various RL settings showcase the remarkable performance of SCRN compared to first-order methods.
Subjects: Machine Learning (cs.LG); Optimization and Control (math.OC)
Cite as: arXiv:2205.12856 [cs.LG]
  (or arXiv:2205.12856v1 [cs.LG] for this version)

Submission history

From: Saber Salehkaleybar [view email]
[v1] Wed, 25 May 2022 15:33:00 GMT (6733kb,D)

Link back to: arXiv, form interface, contact.