The research objective of this award is to develop machine learning techniques to enhance the performance of mobile radiation detection systems. The research will produce methods for detection and explanation of complex patterns in data collected by such systems, while taking into account many more sources of information and factors than a human analyst can process. The developed algorithms will also have the ability to adapt to changing conditions and emerging patterns of threats. The methods will be tested in real-world scenarios against data collected in environments that are awash in ambient, varying background radiation, and which include natural and manufactured radioactive sources that pose no security threat. Deliverables include descriptions of developed algorithms, documentation of the results of their empirical evaluation, software implementations, publications and tutorials.
The results of this research will complement recent developments in detector technology with new ways to understand the data to produce a leap forward in the practice of nuclear detection. The resulting algorithms will enable systems consisting of supervised classification to determine the nature of each detected radiation source, anomaly detection to identify new or unknown sources, identification of groups of self-similar anomalies to discover new classes of sources, and active learning to guide the discovery of new classes and follow-ups on potential threat detections. Jointly, they will allow for increased sensitivity and specificity of nuclear detection system, while containing false detection rates. The results, including evaluations against high-quality data from field measurements and threat simulations conducted in realistic test environments, will be disseminated to enable creation of production-grade systems. The majority of the funding is to support graduate students who will perform the bulk of the research, and who will finish their degrees in 2012-2014 to become new leaders in the multidisciplinary area of machine-learning-supported nuclear threat detection.
