Information Fusion from Unaligned Networks - Research Topic Area 1ci(2)
Abstract
The ability to rapidly and continuously integrate all domains of warfare in multi-domain operations is key to deterring and fighting against near-peer adversaries. Such integration requires the fusion and processing of data from diverse sources to extract critical and actionable intelligence. In particular, many data sources are relational in nature and can be represented as networks, e.g., interconnected sensors, communication between soldiers, networked manned and unmanned platforms, and human-machine collaboration in tactical scenarios. Hence, in order to aid the decision-making of human and artificial intelligence (AI) agents, one needs to combine information from multiple networks and make it available at the right time and at the right place. The main challenge in combining information from different networked systems is the lack of alignment between them, i.e., theagents forming the different networks cannot be directly mapped to each other. Unlike two social networks defined on the sameset of people, it is non-trivial to combine the information of two sensor networks defined on different sensors deployed on different portions of the battlefield. In this context, the objective of this proposal is to derive a comprehensive theory and efficient algorithms to jointly learn from unaligned networks. The vision is to further equip the data-driven and network-centric Army of the future by enabling fast and autonomous extraction of actionable knowledge across all networked systems. The proposed research is centered around three thrusts that broadly aim at: -enhanced identification of networks from partial knowledge by leveraging intelligence obtained from other related networks; -theoretical guarantees and practical considerations of transferring an AI agent trained in one network to another one; and -the generation of novel networks from examples of previous network deployments. The key intellectual novelty enabling the above thrusts is the definition of informative and mathematically tractable ways ofcomparing and combining information from networks without any natural alignment. This is achieved by considering random graph models that can accommodate networks of different sizes, projections of networks into distributions in the space of graph motifs, and graph descriptors that capture structurally relevant information. The impact of the proposed methodologies is expected to be widespread: multimodal data acquisition systems with disparate sensors and heterogeneous information sources are increasingly prevalent. The integration and joint consideration of these data will lead to enhanced warfighters situationalawareness, decision making, command and control, and performance, ultimately redounding in improved national security. At a broader scale, advances in the foundations of learning on unaligned networks will benefit a wide range of areas including robustand adversarial machine learning (ML), graphical modeling, and common applications of graph-based ML such as drug discovery, protein interface prediction, recommender systems, and traffic forecasting.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 19, 2023
- Source ID
- W911NF2310040
Entities
People
- Santiago Segarra
Organizations
- Army Contracting Command
- Rice University
- United States Army