What is Jensen-Shannon divergence and why is it preferred in evaluation?

Jensen-Shannon divergence is a symmetric and bounded version of KL divergence.

JSD is defined as the average KL divergence between each distribution and the mean distribution. It resolves KL’s asymmetry and instability issues, making it more suitable for model monitoring and embedding drift detection. It is always finite and bounded between 0 and 1 (log base 2), which makes interpretation easier in production systems.