On the Interpretability of Regularisation for Neural Networks Through Model Gradient Similarity