The spectacular advances of deep learning in recent years have led to breakthroughs across a wide range of data analysis tasks. At the same time, the techniques used to boost model performance have become increasingly intricate, with ever larger parameter counts and training datasets. What, then, are the practical steps needed to harness these methods for fundamental science especially in the analysis of physics experiments? In this talk we first survey how machine learning and deep learning are being employed in international experimental collaborations. We then present, step by step, the workflow we have developed to tailor deep-learning models for the microscopic image analysis of nuclear emulsion detectors. Finally, we discuss several modern deep-learning ideas and how they might further benefit experimental data analysis.