Impact Of Safeguards Measurement Errors On Deep Neural Networks

The continued growth of the nuclear industry places increasing demands on the International Atomic Energy Agency (IAEA) prompting efforts to improve existing safeguards approaches. There has been considerable recent interest in applying machine learning approaches to safeguards problems given their impressive successes in other domains. However, safeguards data is often accompanied with challenges not present in conventional problems considered in the machine learning literature. One such issue is the presence of random and systematic errors for many measured quantities. This work discusses the impact of the standard safeguards error measurement model on deep neural network models. Exemplar problems are presented and compared with traditional safeguards methods. Results suggest that neural models cannot outperform established statistical approaches using standard neural training methods and loss functions.