Examining the internal representations (activations) of a neural network to understand what it has learned.