A method to identify and visualize interpretable patterns in a neural network's latent representations that drive its predictions.