thank you, that makes sense. Now it's time to really read the article to understand if the difference is the training data or the network topology to be different (although I lean towards the latter).