Different Training performance in release 2018 and 2021

Hello community, based on the example demonstrated in Mathworks for speech denoising using fully convulutional nets I decided to do some research. The problem is, as I tried to train the nets in Matlab 2021 instead of 2018 I got a different performance regarding the RMSE. The release 2018 is doing better. The training set and test set was the same. Can someone please help me understand why? Is there any other default settings that are used in Matlab 2021?

Answer 1

[David Willingham]

Hi Mohamed,

Are you referring to this example?

https://www.mathworks.com/help/deeplearning/ug/denoise-speech-using-deep-learning-networks.html

David

[Mohamed Aserar]

Hi David,

Thank you for the reply. Yes exactly

[jibrahim]

Hi Mohamed,

Are you seeing reduced accuracy? A higher loss? Or is training slower? Which kind of degredation do you see?