THIS GRANT IS A CONTINUATION OF N00014-12-1-0838 Machine Reasoning and Intelligence for Naval Sensing
Abstract
Statement of Work:Proposer will conduct a research programthat has a statistical underpinning, enabling automated systems to provide multiple hypotheses that are (i) consistent with a mission; (ii) support the use of data that is uncertain, incomplete, imprecise, and contradictory (UIIC); (iii) provide a capability to suggest experiments or courses of action that disambiguate between hypotheses; (iv) identify data with appropriate data quality; and (v) represent UIIC data and support efficient computation as well as hypothesis formulation. In particular they will develop graphical methods for timeseries data, data fusion methods combining with Bayesian dictionaries, Nonparametric-Bayesian analysis of heterogeneous space-time data, and Incorporation of graphical priors in Bayesian analysis. They will if warranted expand the work to include Variational robust PCA using improved regularization with the help of beta processes, Combine variational robust PCA with beta processes using an iterative procedure, Temporal modeling of timeseries data embedded in high dimensional graphs, and Concept Drift and Model Refinement Over TimeObjective:Conduct research and provide a statistical underpinning, enabling automatedsystems to provide multiple hypotheses that are (i) consistent with a mission; (ii) support the use of data that is uncertain, incomplete, imprecise, and contradictory (UIIC); (iii) provide a capability to suggest experiments or courses of action that disambiguate between hypotheses; (iv) identify data with appropriate data quality; and (v) represent UIIC data and support efficient computation as well as hypothesis formulation.Approach:The proposed team will integrate distinct and complementary tools to address the challenge of machine reasoning and intelligence. Profs. Bertozzi and Carin will analyze unconventional sources of data, including space-time patterns of human behavior, unstructured data, and HUMINT, while integrating such with traditional sensor data. The graphical constructs developed by Prof. Bertozzi will be integrated as nonparametric priors within the Bayesian formalisms of Prof. Carin. The latter methods will also quantify the value of information, of importance for defining which new data should be acquired to refine inferences, reduce uncertainty on models, and possibly spawn new models. Prof. Osher will constitute the foundation of the proposed program, as his computational tools will make the proposed statistical algorithms tractable for accurate machine reasoning. The Bregman optimization approaches will be integrated within the statistical models developed by Profs. Bertozzi and Carin. Performers will also seek to connect optimization and Bayesian approaches, with variational methods playing an important role. The unification of statistical, Bayesian and optimization methods will be a fundamental product of the proposed program. Additionally, Prof. Osher will develop new techniques for inferring the presence of anomalies in general space-time- spectral data, extending ideas in robustPCA.Overall Merit and ONR Mission/Relevance:The research being conducted directly supports machine reasoning and intelligence which is a key enabler of autononmy and data to decisions which in turn are key pillars for Information Dominance.Progress Statement:The PIs (Osher, Bertozzi, Carin) are developing rigorous computational methods for analyzing and extracting useful information from multimodal data. Osher???s group has developed a method for collecting data for which the last squares estimates for the ranking problemwas maximal Fisher information. They used spectral clustering methods to identify highly-connected communities. Additionally they recovered sparse signals from noisy linear measurements by solving nonlinear differential inclusions. This gives a bias-free and sign-consistent point on the solution paths, a significantly improvement over the LASSO regularization path. They also investigated constrained L1-L2 m
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 26, 2018
- Source ID
- N000141612119
Entities
People
- Stanley Osher
Organizations
- Office of Naval Research
- United States Navy
- University of California, Los Angeles