Artur Dubrawski - Learning to Identify Nuclear Threat

Machine Learning allows making sense of many more sources of information and factors than human analysts possibly can. This talk demonstrates its utility in increasing sensitivity and specificity of detecting nuclear threat while mitigating false alerts. To support these claims, we review results of a project that focuses on spatio-temporal aggregation of evidence collected with mobile gamma-ray spectrometers. Our foundational approach, Bayesian Aggregation, is designed to detect and to spatially localize faint sources of radiation mixed with complex background typical to cluttered urban scenes. It allows inferences in multifactor hypothesis spaces to account for real-world conditions such as noisy and varying background, kinematics of sensor movement, anisotropy and varied intensity of potentially moving sources, and multiplicity of classes of sources. Time permitting, we will also briefly review a subset of other topics of machine learning research being conducted at Carnegie Mellon University Auton Lab that have already made impact outside of academia, and that should yield systems of societal importance in the nearest future.