Batch normalization can cause issues when fine-tuning a pretrained model on a small dataset. What is the typical mitigation?