Real-Time Uncertainty Estimation in Bayesian Deep Learning
Abstract
Deep learning systems are already in wide use in industry in applications such as assisted-driving and medical diagnosis, putting AI safety and robustness to the test on a daily basis. For AI systems to be deployed successfully and responsibly there needs to be reliable and robust AI tools to report when the AI system is about to fail in real-world settings. A pragmatic approach to meet this need often used by industry is to use of Bayesian deep learning (BDL). Gaussian Processes, another approach to this problem, allow for analytic close-form expressions of uncertainty but are difficult to scale. In this work we will develop new deep learning uncertainty tools with analytic close-form expressions which can produce real-time uncertainty estimates. These advances will be achieved by developing theoretical foundations for a recently proposed ad-hoc approach which demonstrated state-of-the-art performance for real-time uncertainty estimation.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2023
- Source ID
- FA86552117017
Entities
People
- Yarin Gal
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Oxford