Probabilistic learning of structure in complex data
Abstract
It has become commonplace in many modern applied domains to collect high-dimensional and complex data, with limited examples of simi"lar data available. This is certainly true in Navy applications as well as in scientific and industry settings. There is hence a cri"tical need to design methods that can learn important structure in big and complex data, efficiently leveraging on any available out""side information, which may take the form of prior experience of individuals familiar with similar data sets, constraints on paramet""ers, and mechanistic information from physical models. In addition, it is crucial to be able to accurately quantify uncertainty in s"tructure learning and in corresponding inferences and predictions based on the available data. This project develops transformative and general new tools to addressthese critically important goals. The proposed methods will significantly outperform existing state-of-the-art methods in settings with limited training data and when significant prior information is available. Particularly novel and potentially influential ideas include the development of methods for (i) much more parsimoniously characterizing low-dimensiona"lstructure, using curved instead of flat component pieces; (ii) more easily and efficiently incorporating known constraints in anal"yses; and (iii) leveraging on mechanistic information available from physical models of the data based on systems of differential eq"uations. Each of these developments will be very broad, accommodating widely different types of data and contexts, making the potent"ial impact highly significant. The public availability of articles describing the methodology and corresponding documented code with worked examples will further increase the impact and enhance the transition to routine use.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 01, 2017
- Source ID
- N000141712844
Entities
People
- David B. Dunson
Organizations
- Duke University
- Office of Naval Research
- United States Navy