A training method that improves generalization by finding parameter values that are insensitive to small perturbations.