What problems arise if batch normalization updates statistics during validation?

If batch normalization updates the statistics (mean and variance) during validation, it will lead to incorrect normalization parameters on validation data. As a result:

• The model will use statistics dependent on the current validation batch, violating the independence principle of validation.
• Validation performance may become unstable and not reflect the true generalization ability.
• It may cause overfitting, as the model adapts to specific validation batches.

Therefore, during validation, it is common to fix the accumulated statistics from training and not update them, to ensure correct model evaluation.