Accession Number:

AD1146054

Title:

Learning Data Representations via Nonconvex Optimization

Descriptive Note:

[Technical Report, Final Report]

Corporate Author:

UNIVERSITY OF SOUTHERN CALIFORNIA

Personal Author(s):

Report Date:

2021-08-17

Pagination or Media Count:

9

Abstract:

In this project we developed a unified understanding of how to design and analyze efficient nonconvex optimization algorithms aimed at learning interpretable representations from data. These data representations can in turn enable automatic knowledge extraction from observed low-level sensory data enhancing a variety of applications. Our main results during the second year can be summarized in three categories 1 understanding the optimization landscape of data representation tasks such as matrix factorization and shallow neural network training, 2 developing principled approaches to utilizing prior knowledge in data representation tasks and characterizing the reduction in the size of the training data that results from such usage of prior knowledge, and 3 developing algorithmic variations that can be implemented on often unreliable modern cloud infrastructure.

Subject Categories:

  • Operations Research

Distribution Statement:

[A, Approved For Public Release]