Universal Multivariate Information Measures for Multisensor Inference (UMIMMI)

The team of Boston Fusion Corp., University of Illinois at Urbana-Champaign, and Syracuse University proposes Phase II of UMIMMI: Universal Multivariate Information Measures for Multisensor Inference, a research program that studies multivariate information measures and uses these to construct optimal, universal algorithms and to lay a framework for general multivariate reasoning. We can then apply this framework to real-world data, particularly data of relevance to the Air Forces challenge of Automatic Target Recognition. In Phase I of UMIMMI, we studied several information measures including two that we defined ourselves; further, we used these measures to develop optimal, universal algorithms for clustering, dimensionality reduction, classification, image registration, and crowdsourcing. These algorithms were published in peer-reviewed journals; we also applied several of them to real-world imagery data, seeing improved accuracy and runtime compared to state-of-the-art universal methods. In Phase II, we propose extending our algorithms to address the new challenges of task difficulty, simultaneous dimensionality reduction and classification, and embedding multivariate distance measures in a reduced dimensionality. Further, we will apply our methods to simulated or AFRL-provided data to explicitly address the challenge of Automatic Target Recognition; we expect that UMIMMI will supplement existing algorithms to provide increased robustness to new threats.