Understanding CNN Hidden Neuron Activations using Concept Induction over Background Knowledge