Improved Regularization and Robustness for Fine-tuning in Neural Networks