The classification of archaeological finds and their representation are shaped by various object epistemological approaches and changes of medium. With ever increasing digitisation, there are now new possibilities of classification, for example using methods of automatic image recognition, as well as the representation of finds on the web with linked open data. ClaReNet, a cooperative project of the Römisch-Germanische Kommission (German Archaeological Institute) and the Big Data Lab (Goethe University Frankfurt), funded by the Bundesministerium für Bildung und Forschung (BMBF; Federal Ministry of Education and Research), tests the possibilities and limits of new digital methods of classification and representation. To this end, traditional approaches of typification and the recording of attributes in numismatics and archaeology are compared with IT-based methods of classification, including deep learning, using the examples of three Celtic coin series that were each chosen to address particular research questions and problems. This work is accompanied by a science and technology study (STS) PANDA, which focuses on Path dependencies, Actor Networks and Digital Agency. This paper briefly introduces the approach of object epistemologies before considering Celtic coins as scientific objects and the history of research on them with regard to classifications and typologies. Using the example of a series of coins from Armorica (Brittany), we will present how deep learning classifies the coinage, how this may differ from classifications by numismatists, and the lessons that are to be learned from this exercise. From an STS perspective, we analyse the actor network that emerges during image data processing. The paper concludes with a reflection on the transformation of numismatic practices resulting from the IT-based methods used in the project, as well as an outlook on further possibilities for research into the classification and representation of Celtic coins.