Learning Data Representations via Nonconvex Optimization
Abstract
It is proposed to develop a unified research effort to design and analyze efficient non-convexoptimization algorithms aimed at learning interpretable representations from copious amounts ofdata. These data representations will enable automatic knowledge extraction from observed low-level sensory data, enhancing a myriad of military applications ranging from target identificationto online surveillance. This project will have three salient features. First, this project provides acomprehensive understanding of when non-convex algorithms yield reliable and globally optimal data representations. This effort will show that a wide variety of interpretable latent representationscan be found from limited and noisy data observations. Second, the PI will characterize thevarious tradeoffs involved between computational and statistical resources, as well as how priorknowledge is incorporated in such algorithms. Finally, the PI will design a range of practically usefulimplementations that are time and space efficient under a streaming setting, enabling solutions toflarge-scale data representation problems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 09, 2018
- Source ID
- FA95501810078
Entities
People
- Mahdi Soltanolkotabi
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Southern California