Nonconvex Information Processing for Heterogeneous and Distributed Data
Abstract
The proposed research program addresses the pressing demand of extracting critical information and actionable intelligence from heterogeneous and decentralized data in a computationally efficient, robust, and provably accurate manner. Modern sensing and imaging technologies produce vast amounts of data at an unprecedented rate. Such data are often generated by highly nonlinear systems, and are aggregated from multiple sensing modalities that might be decentralized at different locations but carry relevant information regarding shared sources of intelligence. The ability to extract such information in a timely, reliable, and low-complexity manner is of great importance to the Army. Inspired by the recent success in nonconvex statistical estimation, this research program aims to develop a comprehensive framework for nonconvex information processing, with an emphasis on large-scale, heterogeneous and distributed data. The aim is to design efficient, robust and accurate optimization-based algorithms for information processing using their natural nonconvex formulations, without resorting to expensive convex relaxations. This is achieved by exploiting the fact that, in concrete signal and information processing scenarios, the nonconvex problems of interest are not generic ones, but armed with much richer structures empowered by statistical modeling of data collection. The proposed program will pursue a deeper understanding of geometric properties of nonconvex loss surfaces and optimization trajectories in the context of statistical estimation, with a focus on more realistic data models that can be heterogeneous and distributed. Specifically, the proposed research program aims at three major thrusts. Thrust 1: Implicit Regularization in Nonconvex Information Processing. Thrust 2: Nonconvex Information Processing for Heterogeneous Data. Thrust 3: Nonconvex Information Processing for Distributed Data. The proposed effort will have far-reaching implications in a broad range of Army applications (e.g. data analytics for Internet-of-Battlefield Things, information fusion and adaptation for networked sensing systems), where it is desirable to extract shared source of critical information from heterogeneous and distributed data in order to enhance the warfightersÕ situation awareness, decision making, command and control, and weapons systems performance. Technology transfers of the proposed methodologies for Army-related missions will be pursued by seeking collaborations with Army Research Lab.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810303
Entities
People
- Yuxin Chen
Organizations
- Army Contracting Command
- Princeton University
- United States Army