Uncertainty-aware Artificial Intelligence for More Effective Decision Making

Abstract This work addresses the determination of uncertainty in AI&ML reasoning when the training data is sparse. Such reasoning should output low uncertainty when the observational sample under test are representative of the training data, but the uncertainty must be high when the test sample is not representative. The characterization of uncertainty is established for low level (i.e., neural networks) and high level (i.e., probabilistic graphical models) reasoning systems.
Authors
  • Lance Kaplan (ARL)
  • Jin-Hee Cho
  • Murat Sensoy
  • Feng Chen
  • Paul Sullivan (ARL)
Date Aug-2018
Venue Army Science & Technology Symposium & Showcase