Robust, Decentralized Feature Learning from Big Data
Abstract
Intelligence, surveillance, target acquisition, and reconnaissance (ISTAR) data is an integral component of effective decision making by the United States Army. The need for a decentralized computational framework that can process big, distributed (and possibly streaming) 1STAR data corrupted in myriad, unexpected ways for robust decision making in dynamic and constrained environments is recognized across the Army leadership. But there is a big gulf between the needs of the US Army in the context of future 1STAR capabilities and the big data research currently being carried out to address the needs of the technology sector (e.g., Google, Facebook, and Twitter). This project develops, analyzes, and validates a computational framework for robust, decentralized processing of ISTAR data that formally takes into account the challenges that are explicitly aligned with the future needs of the Army. The fundamental problems that will be addressed in the project range from robust, decentralized processing algorithms that exploit geographically distributed and contaminated big data for near-optimal inference to efficient techniques for inference from streaming data and message passing strategies that can manage Byzantine failures of some of the distributed entities. The analytical tools that will be leveraged to achieve these objectives include distributed message passing (including distributed consensus), robust statistics, stochastic approximation, and optimization and perturbation theory.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1710546
Entities
People
- Waheed U. Bajwa
Organizations
- Army Contracting Command
- Rutgers University
- United States Army