13.1 Strong Regularisation
Parameterised quantum circuits trained as classifiers face the same challenge as classical models: the need to generalise well to unseen data points. Classically, we have a wide range of supervised learning models and regularisation techniques to choose from. These regularisation techniques that fight overfitting are model specific. For example, we can try to restrict the depth of the decision trees or to impose a penalty term in the cost function when training neural networks.
Consider a conventional feedforward neural network as, arguably, the most direct classical counterpart of a quantum classifier. In both classical and quantum cases, the signal travels through the network in one direction and the layers of quantum gates can be compared to the layers of classical activation units. Regardless of whether we apply L1 (Lasso) or L2 (Ridge) penalty terms, or use dropout techniques, we would like to have a measure of regularisation present in the network...
