On Newton's Method to Unlearn Neural Networks