On Local Overfitting and Forgetting in Deep Neural Networks