I have made a small experiment to convert each network layer gradient norm into a tone in order to synthesize sound from it during training, there are some samples here: http://blog.christianperone.com/2019/08/listening-to-the-neural-network-gradient-norms-during-training/, if someone is interested.