Repairing Neural Networks by Leaving the Right Past Behind