In a machine learning exercice done with
(Tensorflow 2.13.1, Keras 2.13.1) and using binary cross entropy,
I obtain the following loss curve for the training and validation part.

We see some small fluctuations.

Doing exactly the same for more recent version of tensorflow and keras, it starts to be smooth for version >=(Tensorflow 2.16.0rc0, Keras 3.0.0)

Would there be an explanation ?

Formulated in a different way, why is loss starting to be smooth from version >=(Tensorflow 2.16.0rc0, Keras 3.0.0) as compared to old version (Tensorflow 2.13.1, Keras 2.13.1) ?