Abstract
The exceptional performance of machine learning methods has led to their adaptation in many different domains. In the nuclear industry, it has been proposed that machine learning methods have the potential to revolutionize nuclear safety and radiation detection by leveraging that they can be used to augment human and device capabilities. While many applications focus on the accuracy of the learning algorithm's prediction, it has been shown in practice that these algorithms are prone to learn characteristics that are not descriptive or relevant. Hence, this paper focuses on understanding the reasoning behind the classification using saliency methods. Visual representations of the network's learned regions of interest are used to demonstrate whether domain-specific characteristics are being identified, which allows for the end-user to evaluate the performance based on domain knowledge. The results obtained show that focusing on a human-centered approach will ultimately enhance the transparency and trust of the deep learning algorithm's decision.