Deep Learning Discovery Algorithms in High-Energy Physics and the Conflict Between Model-Independence and Robustness

Machine Learning models have become widely recognized for their potential to revolutionize science. In High-Energy Physics, they are of particular promise in propelling the field forward by fueling experimental discoveries. State-of-the-art uses of Machine Learning are considered ‘model-independent’ by the High-Energy Physics community; a verdict shared by several philosophers. In this paper, we dig into the details of unsupervised anomaly detection and show that even modest senses of model-independence are doubt-worthy, once considerations of robustness are factored in. We also offer reasons to think that the theme generalizes beyond the narrow context of High-Energy Physics.