Peering inside the black box by learning the relevance of many-body functions in neural network potentials