Regularisation

A flexible model can fit the noise in its training data. This topic restrains it with a penalty on the size of the weights (ridge), a penalty that sets some weights to zero (the lasso), and by stopping gradient descent early.