We adress the problem of learning concepts in description logics at large scale in a supervised setting. This form of machine learning has a plethora of advantages. First, complex models can be learned from small training data sets by exploiting background knowledge. Moreover, the models computed in this manner are ante-hoc explainable. In this talk, we focus on approchaes which improve the runtime of concept learning approcheas by exploiting multiple representations. For example, we show how binary tensors can be used to improve graph storage and accelerate instance checks. We discuss current challenges daced by this family of approaches as well as some surprising experimental outcomes.