Large-Scale Information Fusion from Multiple Modalities
Abstract
With the rapid advances in processor, storage, and sensing technologies, data available in operational scenarios are becoming increasingly complex. In this context, ~complexity~ means 1) data complexity where variety of data objects/entities are involved, and 2) relational complexity, where complex relationships among data objects and entities are available. The data and relational complexities have provided a unique opportunity to utilize machine intelligence to discover useful information by fusing data sources to make more informed and comprehensive decisions. This isespecially attractive and necessary when we encounter critical decision making situations as in naval operations. But data complexity has also imposed significant challenges on existing machine intelligence algorithms in terms of their effectiveness and efficiency.Proposed Approach. We propose a tensor network based framework to fuse large-scale multi-modal data. In a tensor network, the nodes correspond to individual objects and the edges indicate specific relational structures among them. If we treat the whole data set to be analyzed as a system (which could be represented as a high order tensor), then a tensor network can define a set of rules to approximate tensors with a set of subsystems. Each subsystem can characterize the original data from a certain view point. The connections encode the ~context~ around the subsystems. We will systematically study the possible structures of such a tensor network as well as their construction processes, including both effective learning problem formulations as well as efficient implementations. We will evaluate the proposed framework on publicly available datasets in the domains of medical informatics and multi-modal face recognition that will be generalizable to naval operations. In these domains, decision making will also benefit from joint analysis of multiple data entities andtheir complex relationships. We will develop efficient distributed optimization algorithms that enable large-scale tensor network analysis.Future Naval Relevance. This research is well aligned with the Navy~s mission. One of the key naval capabilities is to be able to automatically make robust decisions on time- critical problems, such as object detection, route planning, hazard and risk estimation, and strategy selection, from a variety of intelligence data types available. To meet this capability, there is an urgent need for developing effective and efficient analytic tools to fuse information from various modalities. The proposed information fusion framework provides integrated solution for large-scale informationfusion from multiple modalities.Significance. The proposed research is significant in following ways: (i) A principled framework to fuse information from multiple modalities to develop models with significantly improved predictive performance. (ii) A distributed optimization algorithm to ensure efficient inference over large-scale distributed data. (iii) An approach applicable to a broad range of applications, not only for naval intelligence analysis, but also medical informatics, social network analysis, and recommender systems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2017
- Source ID
- N000141712265
Entities
People
- Jain Anil
Organizations
- Michigan State University
- Office of Naval Research
- United States Navy