THIS IS A CONTINUATION OF N00014-14-1-0722 Improving the Robustness of Deployed Machine Learning Algorithms
Abstract
The performer will develop theory and algorithms that detect and track model mismatch. They will further develop theory and algorithms that minimize the need for active exemplar labeling when model mismatch is detected. They will develop theory and algorithms for cost functions that lead to maximally invariant low-dimensional representation and reduce the computational complexity of the algorithms. The performer will develop methods to characterize the change in performance of an algorithm that is deployed by measuring differences between the training and test distributions. The performer intends to show that the test error rate can be bounded by the Friedman-Rafsky distance between the training and the test distribution. Furthermore, there is reason to believe that there exists a non-parametric estimator that efficiently estimates this metric. Using this as a baseline, they intend to develop theory and algorithms that can provide a method for automatically detecting model mismatch in deployed algorithms and the environment in which they are operating; develop low-dimensional representations of the data that are maximally invariant to expected changes in the data; and develop methods foridentifying a limited number of test data exemplars for which to request a label (active learning).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 03, 2016
- Source ID
- N000141612156
Entities
People
- Visar Berisha
Organizations
- Arizona State University
- Office of Naval Research
- United States Navy