From Optimization to Sampling for Uncertainty Assessment
Abstract
The goal of this research is to develop efficient methods for assessing uncertainty and sensitivity in high-dimensional models for prediction. Recent years have witnessed tremendous advances in machine learning techniques for building predictive models, but there is currently relatively less understanding of the break-down modes, sensitivity and robustness properties. One way inwhich to address this issues is using efficient algorithms to explore the shape of the posterior distribution produced by predictive models, and to draw samples from this predictive distribution. The research will leverage results from optimization theory and stochastic differential equations, along with associated techniques from Markov chain theory for analyzing mixing times. Expected outcomes of the research include the development of algorithms forexploring posterior distributions, analysis of their mixing times and dependence on dimension and other problem parameters, and implementation of these algorithms, and study of their properties in various domains.Many DoD-related problems can be addressed with machine learning methods for highdimensional prediction and classification, with examples including classification in vehicles in aerial imagery, and anomaly detection in video and other signal streams. The methods developed in this research project will lead to principled ways of assessing the uncertainty in such predictions. Such uncertainty assessment is essential if the output of a predictive model areto be fed into different components of a larger system.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 27, 2018
- Source ID
- N000141812640
Entities
People
- Martin J. Wainwright
Organizations
- Office of Naval Research
- United States Navy
- University of California Regents