A Mirror Descent Perspective of Smoothed Sign Descent

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wang, Shuyang, Klabjan, Diego
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914978136588288
author Wang, Shuyang
Klabjan, Diego
author_facet Wang, Shuyang
Klabjan, Diego
contents Recent work by Woodworth et al. (2020) shows that the optimization dynamics of gradient descent for overparameterized problems can be viewed as low-dimensional dual dynamics induced by a mirror map, explaining the implicit regularization phenomenon from the mirror descent perspective. However, the methodology does not apply to algorithms where update directions deviate from true gradients, such as ADAM. We use the mirror descent framework to study the dynamics of smoothed sign descent with a stability constant $\varepsilon$ for regression problems. We propose a mirror map that establishes equivalence to dual dynamics under some assumptions. By studying dual dynamics, we characterize the convergent solution as an approximate KKT point of minimizing a Bregman divergence style function, and show the benefit of tuning the stability constant $\varepsilon$ to reduce the KKT error.
format Preprint
id arxiv_https___arxiv_org_abs_2410_14158
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle A Mirror Descent Perspective of Smoothed Sign Descent
Wang, Shuyang
Klabjan, Diego
Machine Learning
Optimization and Control
Recent work by Woodworth et al. (2020) shows that the optimization dynamics of gradient descent for overparameterized problems can be viewed as low-dimensional dual dynamics induced by a mirror map, explaining the implicit regularization phenomenon from the mirror descent perspective. However, the methodology does not apply to algorithms where update directions deviate from true gradients, such as ADAM. We use the mirror descent framework to study the dynamics of smoothed sign descent with a stability constant $\varepsilon$ for regression problems. We propose a mirror map that establishes equivalence to dual dynamics under some assumptions. By studying dual dynamics, we characterize the convergent solution as an approximate KKT point of minimizing a Bregman divergence style function, and show the benefit of tuning the stability constant $\varepsilon$ to reduce the KKT error.
title A Mirror Descent Perspective of Smoothed Sign Descent
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2410.14158