Angie Sheffield, Senior Program Manager, Data Science
National Nuclear Security Administration
Office of Defense Nuclear Nonproliferation Research and Development (NA-22)
Domain-aware artificial intelligence (AI) is an area of keen interest for DNN R&D. Domain-aware methods are key to the development of next-generation AI technologies suitable for the unique challenges of nuclear proliferation detection and national security.
Conventional machine and deep learning (ML/DL) models and approaches are insufficient for these challenges. In fact, nuclear proliferation detection is a nearly intractable science and engineering challenge for any technique or discipline. With the development and use of domain-aware AI techniques, which leverage the power of AI and the power of our knowledge, we have a chance.
Domain-aware methods play to our strengths. They require innovative and multidisciplinary approaches and draw on the nuclear weapons expertise and longstanding computational capabilities of the National Laboratories and our academic partners.
This workshop will focus on the use of domain-aware methods to develop AI systems that enhance U.S. nuclear proliferation detection and address the unique challenges of this mission:
Complex and noisy environments: Domain-aware AI techniques to combine heterogeneous data sources, modeled predictions, and ML may increase sensitivity to faint signals of interest. Domain-informed techniques that train ML to model source and propagation characteristics may produce an entirely new class of detection methods that no longer rely on minimizing signal-to-noise.
Sparse data and rare events: Events of interest are rare, and assumptions of independent and uniformly distributed samples do not hold. Methods to combine expert knowledge, models of the nuclear weapons development process, and ML may be used to augment sparse samples or provide underlying structure that cannot be learned directly from the data.
Robust deployment and decision support: Domain-aware methods may be used to validate that an AI model accurately predicts system behavior, not just fits the data. Model performance must be domain-informed to assess the limits of an AI system to support decision making and build AI systems that perform predictably in new and uncharacterized settings.
Early proliferation detection and signature discovery: Advances in ML and the availability of new data sources present new opportunities to detect early indicators of nuclear proliferation from large and unstructured data. Domain-aware techniques may help to direct exploitation of these new data sources or interpret the signatures identified by ML models.
This research is incredibly challenging. We know we are at the cutting edge because it is this hard – the easy science has already been done. We know this is the right research because U.S. government partners are using next-generation AI systems we have developed to successfully detect nuclear proliferation.
We are thrilled for your participation in our workshop. We look forward to working together to advance the field of AI and develop capabilities to reduce the threat of nuclear weapons proliferation.
Next Gen AI Workshop Series
The Data Science Research Portfolio within the National Nuclear Security Administration (NNSA) Office of Defense Nuclear Nonproliferation Research and Development (NA-22) is driving development of the next generation in artificial intelligence to advance the state-of-the-art and build AI systems suitable for national security where high-consequence decisions require beyond state-of-the-art AI, as commercial AI is insufficient to solve mission challenges.
The Next Gen AI Workshop Series is focused on proliferation detection mission-driven gaps and requirement define R&D focus to drive the right new math and science across five areas:
- Robust AI Models
- Domain-Aware Methods
- Interpretable Model and Explainability Techniques
- Highly-Specialized AI Systems
- Curating and Generating Relevant Datasets
The first workshop was held on September 15-16, 2020 on Interpretable Model and Explainability Techniques, report is forthcoming in March 2021. A recent article AI research and representation: One team’s mission to pursue both about the workshop reflects our team’s commitment to inclusive diversity. A third workshop is planned for July 2021!
There will be a workshop report summarizing the abstracts and results of the workshop that will be made published and made available to read. We also plan to invite selected presentations to write up longer papers on their work, which we will seek to publish as a stand-alone collection describing progress in next-generation AI R&D related to the issues that are characteristically found in the proliferation detection mission.
Diversity and Inclusion Policy
The Workshop on Advances in Domain Aware AI for Proliferation Detection is committed to driving inclusive diversity and representation across our field. To foster an organizational culture that welcomes all perspectives, meeting organizers will deliberately consider the role of inclusive diversity throughout the planning process and in the recruitment of diverse subject matter experts participating in our event(s). Our goals include improving gender equity and representation, highlighting the work of women and minorities in STEM, and contributing to the Department of Energy’s larger initiative to recruit a talented and diverse workforce.