You can use the same trick to make your models sparse. Less parameters means less flexibility to interpret output "correctly".