Structure-Based Optimization for Data Science

Abstract

General-purpose iterative optimization algorithms such as (stochastic) gradient descent are the workhorse tools in modern data analy,sis. However, as these algorithms have been designed to apply to broad classes of optimization problems in a black-box manner, they,are subject to strict lower bounds that prevent further speedups in the worst case. The goal of this proposal is to design optimizat,ion algorithms that go beyond the worst case predicted by the lower bounds by taking a white-box approach that allows utilizing th,e structure present in common data science problems. Concrete examples of structure that is common in learning tasks and that this p,roject aims to exploit are the separability of common loss functions, min-max perspective, and non-negativity in different problem formulations.

Document Details

Document Type
DoD Grant Award
Publication Date
Apr 01, 2022
Source ID
N000142212348

Entities

People

  • Jelena Diakonikolas

Organizations

  • Office of Naval Research
  • United States Navy
  • University of Wisconsin System

Tags

Fields of Study

  • Computer science

Readers

  • Educational Psychology
  • Neural Network Machine Learning.
  • Operations Research

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms