A model uses batch normalization and is trained with a very small batch size of 2. What problem arises?