Abstract

In recent years, Machine Learning (ML) techniques have become more and more important in high energy physics (HEP). Even more recently, the HEP community has begun to explicitly acknowledge the “black box”-nature of ML algorithms and associated, induced features of opacity (e.g. Chang et al. 2018).
This raises, among others, the following questions: Are these opacities different from the ones which are associated with computer simulation? Are they contingent on the nature of cognitive agents? Does ML introduce non-contingent opacity?
In this talk, I want to distinguish between different kinds of opacity that could be introduced by ML algorithms: (i) opacity induced by complexity, which is a feature of all traditional experimental procedures in modern HEP; (ii) opacity induced by method, which is the specific opacity that accompanies searches using ML techniques. 
In a final step I want to change the focus from the question how the machine is learning to the question what the machine is learning. In this context, I define a third kind of opacity: model-opacity, which describes the opacity that is caused by the underdetermination of the connection between ML models and the underlying physical phenomenon.