Uncertainty Quantification for Machine Learning
Abstract
Machine learning (ML) and artificial intelligence (AI) have made great strides in several application areas, such as computer vision and natural language processing. However these successes usually concern problems with relatively low levels of uncertainty; and uncertainty quantification for machine learning remains a very challenging problem. At the same time, accurate uncertainty quantification could benefit the U.S. Army by improving learning and decision systems under imperfect information, and thus can be a crucial part of the vision described in the Army Modernization Strategy. This project proposes to develop new algorithms for uncertainty quantification in machine learning. These algorithms will have rigorous theoretical guarantees showing that they work well under reasonable conditions. They will also be evaluated on challenging realistic datasets with high levels of uncertainty. The work will build on the PI s experience in uncertainty quantification and statistical machine learning. The proposal is divided into three main thrusts: Efficiently verifying and improving calibration, online prediction sets, and rigorous anomaly detection. Page
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 27, 2023
- Source ID
- W911NF2310296
Entities
People
- Edgar Dobriban
Organizations
- Army Contracting Command
- United States Army
- University of Pennsylvania