Adaptive Exploitation of Non-Commutative Multimodal Information Structure
Abstract
Our interdisciplinary team will develop a new comprehensive information theory for data collection in non-commutative information structures intrinsic to hierarchical representations, distributed sensing, and adaptive online processing. Tools will be developed based on our novel theory in conjunction with the latest theories of information, random matrices, free probability, optimal transport, and statistical machine learning. They will be applied to the technical domains of causal inference, adaptive learning, computer vision, and heterogeneous sensor networks and will be validated on real-data testbeds including: 1) human action and collective behavior recognition; and 2) crowd-sourcing in a network of brain-machine-interfaces (BCI). Our framework will provide answers to questions such as: What are the fundamental performance limits for non-commutative information collection and processing systems? What is the effect of side information on noncommutative information structures? How can low complexity proxies for performance be defined that approximate or bound non-commutative performance limits? How can non-commutativity of adaptive measurements be exploited to improve fusion, processing, and planning for distributed sensing systems? When do sequential or partially ordered designs offer significant performance gains relative to randomized designs, like compressive sensing? Our approaches for extracting knowledge from complex irreversible partially ordered information structures include but are not limited to introduction of information divergence measures over non-commutative algebras, non-commutative relative entropy measures, and estimation techniques for such measures for high-dimensional data. Accounting for non-commutative structures will result in fundamentally new ways of fusing ordered, directed, or hierarchical organized information in order to support timely decisions at the appropriate level of granularity. Humans learn actively and adaptively, and their judgments about the likelihood of events and dependencies among variables are strongly influenced by the perception of cause and effect, where as man-made systems only employ correlation-type symmetric measures of dependencies. We will develop a theory of decentralized information sharing, causal inference, and active learning inspired by human decision making. Establishment of such a theory for sensing and data processing and application of it to grand challenges in computer vision and BCIs will provide new capabilities, including improved time-sensitive, dynamic, multi-source information processing, actuation, and performance prediction guarantees. Research advances in this area will directly impact the problem of discovering, monitoring, and understanding situational dynamics of hostile environments, and hence will significantly advance the efforts of the DoD to increase situational awareness and ensure national security. The team will work closely with ARL and other DoD labs in realizing this mission.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 02, 2019
- Source ID
- W911NF1510479
Entities
People
- Rayadurgam Srikant
Organizations
- Army Contracting Command
- United States Army
- University of Illinois Urbana–Champaign